diff --git a/.git-blame-ignore-revs b/.git-blame-ignore-revs
deleted file mode 100644
index 17d96cd0f5..0000000000
--- a/.git-blame-ignore-revs
+++ /dev/null
@@ -1,8 +0,0 @@
-# Commits listed here are skipped by `git blame` so that bulk, whitespace-only
-# changes don't obscure the real authorship of a line.
-#
-# GitHub honors this file automatically. To use it locally, run once:
-#   git config blame.ignoreRevsFile .git-blame-ignore-revs
-
-# chore(studio/frontend): normalize line endings to LF
-c50b8ab910f5aa56dd7ae0022d2c7b96bfe3384a
diff --git a/.github/workflows/consolidated-tests-ci.yml b/.github/workflows/consolidated-tests-ci.yml
index f7c338d76b..b56a6c2615 100644
--- a/.github/workflows/consolidated-tests-ci.yml
+++ b/.github/workflows/consolidated-tests-ci.yml
@@ -2204,12 +2204,13 @@ jobs:
           pip install -e "$RUNNER_TEMP/unsloth-zoo" --no-deps
           pip show unsloth_zoo
 
-      - name: llama.cpp install via unsloth_zoo.llama_cpp + `llama-cli --help` smoke
+      - name: llama.cpp install via unsloth_zoo.llama_cpp + CLI `--help` smoke
         # Exercise the canonical `unsloth_zoo.llama_cpp.install_llama_cpp`
         # flow that GGUF export uses at runtime: clone ggml-org/llama.cpp
         # into ~/.unsloth/llama.cpp, build the LLAMA_CPP_TARGETS list
         # (llama-quantize, llama-cli, llama-mtmd-cli, llama-gguf-split,
-        # llama-server) via cmake, then run `llama-cli --help`.
+        # llama-server) via cmake, then run `--help` on whichever CLI
+        # inference binary the build actually produced.
         #
         # This replaces the previous "download upstream prebuilt zip"
         # approach, which silently exited 0 with the message
@@ -2218,6 +2219,18 @@ jobs:
         # matched their current asset names). The build path is the same
         # one Unsloth users hit in production via `model.save_pretrained_gguf`.
         #
+        # We do NOT hard-require `llama-cli` specifically: upstream
+        # ggml-org/llama.cpp moved the cli/server/ui targets behind the
+        # `LLAMA_BUILD_SERVER` cmake option (tools/CMakeLists.txt) and the
+        # set of binaries that survive a given checkout drifts over time
+        # (e.g. a recent build root shipped llama-server + llama-quantize
+        # + llama-diffusion-cli but no llama-cli). The durable contract is
+        # "install_llama_cpp produced a working CLI inference binary AND a
+        # working quantizer", so we --help-probe the first of
+        # llama-cli / llama-mtmd-cli / llama-server that exists. If a
+        # future llama.cpp restores llama-cli it is first in the list and
+        # is preferred, so this stays backwards compatible.
+        #
         # Wall-time budget: ~3-5 min cold, dominated by cmake build of
         # 5 targets on the runner's 4 cores. Apt-package install is
         # handled by `install_llama_cpp` itself via its
@@ -2252,8 +2265,9 @@ jobs:
           print(f"Build targets: {LLAMA_CPP_TARGETS}")
           # install_llama_cpp returns (quantizer_path, converter_script_path).
           # The quantizer's directory is the `llama.cpp` install root, which
-          # also holds llama-cli after build/bin/llama-* gets copied up
-          # (llama_cpp.py:867-871).
+          # also holds the CLI inference binaries after build/bin/llama-* gets
+          # copied up (llama_cpp.py:1450-1454; on Windows they stay in
+          # build/bin/Release/).
           quantizer, converter = install_llama_cpp(print_output=True)
           assert quantizer and os.path.exists(quantizer), (
               f"install_llama_cpp returned quantizer={quantizer!r} but file missing"
@@ -2262,25 +2276,54 @@ jobs:
               f"install_llama_cpp returned converter={converter!r} but missing"
           )
           install_root = os.path.dirname(quantizer)
-          cli = os.path.join(install_root, "llama-cli")
-          assert os.path.exists(cli), (
-              f"llama-cli not found at {cli!r} after build. Build root contents: "
-              f"{sorted(p for p in os.listdir(install_root) if p.startswith('llama-'))[:20]}"
-          )
-          assert os.access(cli, os.X_OK), f"{cli!r} not executable"
-          # `llama-cli --help` exits non-zero on some builds; the contract
-          # is that recognizable help text appears on stdout/stderr.
+          is_windows = sys.platform == "win32"
+          exe = ".exe" if is_windows else ""
+          # Search both the copied-up root and the Windows build/bin/Release/
+          # location the quantizer might already live in.
+          search_dirs = [install_root]
+          win_release = os.path.join(install_root, "build", "bin", "Release")
+          if win_release not in search_dirs:
+              search_dirs.append(win_release)
+          # Any of these proves a working llama.cpp CLI inference binary was
+          # built. Order = preference: llama-cli is canonical (restored first
+          # if upstream brings it back), then the multimodal CLI, then the
+          # server (always built whenever cli would be, behind LLAMA_BUILD_SERVER).
+          cli_names = [f"llama-cli{exe}", f"llama-mtmd-cli{exe}", f"llama-server{exe}"]
+          cli = None
+          cli_name = None
+          for name in cli_names:
+              for d in search_dirs:
+                  candidate = os.path.join(d, name)
+                  if os.path.exists(candidate) and (is_windows or os.access(candidate, os.X_OK)):
+                      cli, cli_name = candidate, name
+                      break
+              if cli is not None:
+                  break
+          if cli is None:
+              found = []
+              for d in search_dirs:
+                  if os.path.isdir(d):
+                      found += [p for p in os.listdir(d) if p.startswith("llama-")]
+              raise AssertionError(
+                  f"No CLI inference binary ({', '.join(cli_names)}) found after "
+                  f"build in {search_dirs}. Build root contents: {sorted(set(found))[:20]}"
+              )
+          print(f"Using CLI inference binary: {cli_name} -> {cli}")
+          # `--help` exits non-zero on some builds; the contract is that
+          # recognizable help text appears on stdout/stderr. llama-server
+          # exposes a different flag set than llama-cli, so accept its
+          # tokens too (e.g. --host / --port / "server").
           proc = subprocess.run(
               [cli, "--help"], capture_output=True, text=True, timeout=30,
           )
           combined = (proc.stdout or "") + (proc.stderr or "")
-          print("--- llama-cli --help (first 30 lines) ---")
+          print(f"--- {cli_name} --help (first 30 lines) ---")
           print("\n".join(combined.splitlines()[:30]))
           assert any(
               tok in combined.lower()
-              for tok in ("usage", "--help", "--model", "-m,")
+              for tok in ("usage", "--help", "--model", "-m,", "--host", "--port", "server")
           ), (
-              f"llama-cli --help produced no recognizable help text. "
+              f"{cli_name} --help produced no recognizable help text. "
               f"exit={proc.returncode}\nstdout: {proc.stdout[:400]!r}\n"
               f"stderr: {proc.stderr[:400]!r}"
           )
@@ -2296,7 +2339,7 @@ jobs:
               f"stderr: {q.stderr[:400]!r}"
           )
           print(
-              f"\nOK: install_llama_cpp produced a working llama-cli at {cli} "
+              f"\nOK: install_llama_cpp produced a working {cli_name} at {cli} "
               f"and llama-quantize at {quantizer}."
           )
           PY
diff --git a/.github/workflows/local-agent-guides-ci.yml b/.github/workflows/local-agent-guides-ci.yml
index 150dbc3fde..299ee3f18b 100644
--- a/.github/workflows/local-agent-guides-ci.yml
+++ b/.github/workflows/local-agent-guides-ci.yml
@@ -64,6 +64,12 @@ concurrency:
 permissions:
   contents: read
 
+# Secret handling on pull_request: these jobs check out and run PR-controlled code
+# (install.sh, .github/scripts/**), so HF_TOKEN (an external HF credential) is gated
+# off pull_request at each step below -- public GGUF repos still download anonymously.
+# GH_TOKEN (GITHUB_TOKEN) is kept: it is the job-scoped contents:read token and
+# install_llama_prebuilt.py needs it for the GitHub releases API (else 403s).
+
 env:
   # Determinism precedent (studio-inference-smoke.yml): temp 0 + fixed seed.
   UNSLOTH_SEED: '3407'
@@ -134,7 +140,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -150,7 +157,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -335,7 +343,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -351,7 +360,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -508,7 +518,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -524,7 +535,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Gated off PR (see note above); public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/mlx-ci.yml b/.github/workflows/mlx-ci.yml
index 221e86f235..864630f9f0 100644
--- a/.github/workflows/mlx-ci.yml
+++ b/.github/workflows/mlx-ci.yml
@@ -241,7 +241,8 @@ jobs:
       # non-zero binary exit is an Unsloth/Studio bug.
       - name: Studio prebuilt llama.cpp install + GGUF inference (Mac M1)
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
           # install_llama_prebuilt.py hits the GitHub releases API to
           # resolve the asset URL. Anonymous calls share the runner-IP
           # rate-limit bucket and 403 quickly -- pass the workflow's
@@ -332,7 +333,8 @@ jobs:
       # train_metrics.json so we can detect regressions across CI runs.
       - name: MLX export round-trip — TRAIN + SAVE 3 formats
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
           UNSLOTH_COMPILE_DISABLE: '1'
         run: |
           mkdir -p mlx_workdir
@@ -348,7 +350,8 @@ jobs:
       # the saved dir.
       - name: MLX export round-trip — RELOAD LoRA (fresh process)
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
           UNSLOTH_COMPILE_DISABLE: '1'
         run: |
           python tests/studio/run_real_mlx_smoke.py reload \
@@ -357,7 +360,8 @@ jobs:
 
       - name: MLX export round-trip — RELOAD merged_16bit (fresh process)
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
           UNSLOTH_COMPILE_DISABLE: '1'
         run: |
           python tests/studio/run_real_mlx_smoke.py reload \
@@ -372,7 +376,8 @@ jobs:
       # LoRA + merged_16bit assertions remain the gating signal.
       - name: MLX export round-trip — RELOAD GGUF via llama-cli (fresh process)
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           if python -c "import json,sys; m=json.load(open('mlx_workdir/train_metrics.json')); sys.exit(0 if m.get('gguf_supported') else 1)"; then
             python tests/studio/run_real_mlx_smoke.py reload \
diff --git a/.github/workflows/studio-api-smoke.yml b/.github/workflows/studio-api-smoke.yml
index b196805cf7..15efee382e 100644
--- a/.github/workflows/studio-api-smoke.yml
+++ b/.github/workflows/studio-api-smoke.yml
@@ -83,7 +83,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -100,7 +101,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-backend-ci.yml b/.github/workflows/studio-backend-ci.yml
index 6e53a290cf..ea60252cf6 100644
--- a/.github/workflows/studio-backend-ci.yml
+++ b/.github/workflows/studio-backend-ci.yml
@@ -226,6 +226,7 @@ jobs:
               tests/sh/test_studio_home_node_dir.sh \
               tests/sh/test_system_node_readonly.sh \
               tests/sh/test_nvcc_meets_llama_minimum.sh \
+              tests/sh/test_resolve_cuda_archs.sh \
               tests/sh/test_tauri_install_exit_order.sh \
               tests/sh/test_torch_constraint.sh \
               tests/sh/test_torch_flavor.sh; do
diff --git a/.github/workflows/studio-inference-smoke.yml b/.github/workflows/studio-inference-smoke.yml
index c2c4fa03bf..aebf90380a 100644
--- a/.github/workflows/studio-inference-smoke.yml
+++ b/.github/workflows/studio-inference-smoke.yml
@@ -97,7 +97,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -114,7 +115,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -364,7 +366,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -380,7 +383,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -845,7 +849,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -863,7 +868,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-api-smoke.yml b/.github/workflows/studio-mac-api-smoke.yml
index 412726538c..617ce189dc 100644
--- a/.github/workflows/studio-mac-api-smoke.yml
+++ b/.github/workflows/studio-mac-api-smoke.yml
@@ -68,7 +68,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -85,7 +86,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-inference-smoke.yml b/.github/workflows/studio-mac-inference-smoke.yml
index c794a34acd..d562294d42 100644
--- a/.github/workflows/studio-mac-inference-smoke.yml
+++ b/.github/workflows/studio-mac-inference-smoke.yml
@@ -91,7 +91,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -110,7 +111,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -346,7 +348,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -363,7 +366,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -725,7 +729,8 @@ jobs:
         # Authenticated + parallel: shared macos-14 NAT egress stalls
         # multi-GB anonymous downloads.
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -752,7 +757,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-install-matrix.yml b/.github/workflows/studio-mac-install-matrix.yml
index da944d4b5c..362305cdd4 100644
--- a/.github/workflows/studio-mac-install-matrix.yml
+++ b/.github/workflows/studio-mac-install-matrix.yml
@@ -63,7 +63,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-ui-smoke.yml b/.github/workflows/studio-mac-ui-smoke.yml
index 4f9f94b534..512af54d53 100644
--- a/.github/workflows/studio-mac-ui-smoke.yml
+++ b/.github/workflows/studio-mac-ui-smoke.yml
@@ -68,7 +68,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -85,7 +86,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-mac-update-smoke.yml b/.github/workflows/studio-mac-update-smoke.yml
index f554a16415..d104306c7e 100644
--- a/.github/workflows/studio-mac-update-smoke.yml
+++ b/.github/workflows/studio-mac-update-smoke.yml
@@ -62,7 +62,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -74,7 +75,8 @@ jobs:
       - name: First update should be a no-op (prebuilt already validated)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -93,7 +95,8 @@ jobs:
       - name: Second update must also be a no-op
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log
diff --git a/.github/workflows/studio-ui-smoke.yml b/.github/workflows/studio-ui-smoke.yml
index dcf9fd26af..297a585430 100644
--- a/.github/workflows/studio-ui-smoke.yml
+++ b/.github/workflows/studio-ui-smoke.yml
@@ -82,7 +82,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -99,7 +100,8 @@ jobs:
       - name: Install Studio (--local, --no-torch)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
diff --git a/.github/workflows/studio-update-smoke.yml b/.github/workflows/studio-update-smoke.yml
index 307bb51972..08a79afacd 100644
--- a/.github/workflows/studio-update-smoke.yml
+++ b/.github/workflows/studio-update-smoke.yml
@@ -71,7 +71,8 @@ jobs:
         # prebuilt path falls back to source build.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           mkdir -p logs
           set -o pipefail
@@ -86,7 +87,8 @@ jobs:
         # idempotency regressed.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -109,7 +111,8 @@ jobs:
         # the first one.
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log
diff --git a/.github/workflows/studio-windows-api-smoke.yml b/.github/workflows/studio-windows-api-smoke.yml
index 78efe918ac..e9abd2d669 100644
--- a/.github/workflows/studio-windows-api-smoke.yml
+++ b/.github/workflows/studio-windows-api-smoke.yml
@@ -75,7 +75,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -124,7 +125,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
diff --git a/.github/workflows/studio-windows-inference-smoke.yml b/.github/workflows/studio-windows-inference-smoke.yml
index a997c89b67..c44c68278d 100644
--- a/.github/workflows/studio-windows-inference-smoke.yml
+++ b/.github/workflows/studio-windows-inference-smoke.yml
@@ -127,7 +127,8 @@ jobs:
         # described above (outcome != success).
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -179,7 +180,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -476,7 +478,8 @@ jobs:
         id: download-gguf
         if: steps.cache-gguf.outputs.cache-hit != 'true' || steps.cache-gguf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p gguf-cache
@@ -524,7 +527,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -906,7 +910,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -956,7 +961,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -1299,7 +1305,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -1376,7 +1383,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           $ProgressPreference = 'SilentlyContinue'
diff --git a/.github/workflows/studio-windows-ui-smoke.yml b/.github/workflows/studio-windows-ui-smoke.yml
index 00458d213b..405309916a 100644
--- a/.github/workflows/studio-windows-ui-smoke.yml
+++ b/.github/workflows/studio-windows-ui-smoke.yml
@@ -91,7 +91,8 @@ jobs:
         id: prime-hf
         if: steps.cache-hf.outputs.cache-hit != 'true' || steps.cache-hf.outcome != 'success'
         env:
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           python -m pip install --upgrade huggingface_hub
           mkdir -p hf-cache
@@ -155,7 +156,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 redirects ALL PowerShell streams (stdout, stderr,
diff --git a/.github/workflows/studio-windows-update-smoke.yml b/.github/workflows/studio-windows-update-smoke.yml
index 1a2a7df493..888b3d70a3 100644
--- a/.github/workflows/studio-windows-update-smoke.yml
+++ b/.github/workflows/studio-windows-update-smoke.yml
@@ -133,7 +133,8 @@ jobs:
         shell: pwsh
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           New-Item -ItemType Directory -Force -Path logs | Out-Null
           # *>&1 captures Write-Host (Information stream) output;
@@ -180,7 +181,8 @@ jobs:
       - name: First update should be a no-op (prebuilt already validated)
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update.log
@@ -199,7 +201,8 @@ jobs:
       - name: Second update must also be a no-op
         env:
           GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
-          HF_TOKEN: ${{ secrets.HF_TOKEN }}
+          # Withheld on PR: this step runs checked-out PR code; public GGUF still downloads.
+          HF_TOKEN: ${{ github.event_name != 'pull_request' && secrets.HF_TOKEN || '' }}
         run: |
           set -o pipefail
           unsloth studio update --local 2>&1 | tee logs/update2.log
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index bf2a0c8e7c..8dcb9130b3 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -1,6 +1,6 @@
 repos:
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    rev: v0.15.17
+    rev: v0.15.18
     hooks:
       - id: ruff
         args:
diff --git a/install.ps1 b/install.ps1
index 765f33b1ff..745c7369ea 100644
--- a/install.ps1
+++ b/install.ps1
@@ -2146,7 +2146,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated no-torch)" { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -2160,7 +2160,7 @@ exit 0
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (migrated)" { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -2226,7 +2226,7 @@ exit 0
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (no-torch)" { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommandRetry -Label "install pydantic" { uv pip install --python $VenvPython pydantic }
@@ -2238,7 +2238,7 @@ exit 0
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6" }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (local)" { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7" }
         } else {
             $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth" { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -2266,7 +2266,7 @@ exit 0
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.6.6" "unsloth>=2026.6.8" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommandRetry -Label "install unsloth (auto torch backend)" { uv pip install --python $VenvPython "unsloth-zoo>=2026.6.7" "unsloth>=2026.6.9" --torch-backend=auto }
             if ($baseInstallExit -ne 0) {
                 Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
                 return (Exit-InstallFailure "Failed to install unsloth (exit code $baseInstallExit)" $baseInstallExit)
@@ -2595,6 +2595,7 @@ exit 0
             step "launch" "to start later, run:"
             substep "unsloth studio -p 8888"
             substep "(add -H 0.0.0.0 to allow network / cloud access)"
+            substep "(add --secure to allow HTTPS)"
             Write-Host ""
         }
     } else {
@@ -2615,6 +2616,7 @@ exit 0
             substep "unsloth studio -p 8888"
         }
         substep "(add -H 0.0.0.0 to allow network / cloud access)"
+        substep "(add --secure to allow HTTPS)"
         Write-Host ""
     }
 }
diff --git a/install.sh b/install.sh
index 7a2e18e374..62cc2920e6 100755
--- a/install.sh
+++ b/install.sh
@@ -2621,7 +2621,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd_retry "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6"
+            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
         # Resolve pydantic WITH deps so pip pins pydantic-core to the
         # matching version (no-torch-runtime.txt below is --no-deps).
         # All transitive deps are torch-free.
@@ -2634,7 +2634,7 @@ if [ "$_MIGRATED" = true ]; then
     else
         run_install_cmd_retry "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6"
+            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2838,7 +2838,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd_retry "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6"
+            "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd_retry "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2856,7 +2856,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd_retry "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.6.8" "unsloth-zoo>=2026.6.6"
+            --upgrade-package unsloth "unsloth>=2026.6.9" "unsloth-zoo>=2026.6.7"
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2888,7 +2888,7 @@ else
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.6.6" "unsloth>=2026.6.8" --torch-backend=auto
+        run_install_cmd_retry "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" "unsloth-zoo>=2026.6.7" "unsloth>=2026.6.9" --torch-backend=auto
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -3160,6 +3160,7 @@ if [ -t 1 ]; then
             step "launch" "to start later, run:"
             substep "unsloth studio -p 8888"
             substep "(add -H 0.0.0.0 to allow network / cloud access)"
+            substep "(add --secure to allow HTTPS)"
             echo ""
             ;;
     esac
@@ -3181,5 +3182,6 @@ else
         substep "unsloth studio -p 8888"
     fi
     substep "(add -H 0.0.0.0 to allow network / cloud access)"
+    substep "(add --secure to allow HTTPS)"
     echo ""
 fi
diff --git a/pyproject.toml b/pyproject.toml
index 83d65bc1a3..dd957766b9 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -71,7 +71,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.6.6",
+    "unsloth_zoo>=2026.6.7",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -92,7 +92,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.6.6",
+    "unsloth_zoo>=2026.6.7",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -582,7 +582,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.6.6",
+    "unsloth_zoo>=2026.6.7",
     "packaging",
     "tyro",
     "transformers>=4.51.3,!=4.52.0,!=4.52.1,!=4.52.2,!=4.52.3,!=4.53.0,!=4.54.0,!=4.55.0,!=4.55.1,!=4.57.0,!=4.57.4,!=4.57.5,!=5.0.0,!=5.1.0,<=5.5.0",
diff --git a/scripts/scan_packages_baseline.json b/scripts/scan_packages_baseline.json
index 67952c24f1..f953f4d206 100644
--- a/scripts/scan_packages_baseline.json
+++ b/scripts/scan_packages_baseline.json
@@ -72,13 +72,6 @@
       "severity": "CRITICAL",
       "evidence": "Archive: L317: a['TarFileType'] = tarfile.open(fileobj=_fileW,mode='w')\nNetwork: L330: x['SocketType'] = _socket = socket.socket()"
     },
-    {
-      "package": "evaluate",
-      "file": "evaluate/utils/file_utils.py",
-      "check": "C2 polling/beaconing loop detected",
-      "severity": "CRITICAL",
-      "evidence": "L261: while True:"
-    },
     {
       "package": "execnet",
       "file": "execnet/gateway_base.py",
@@ -401,27 +394,6 @@
       "severity": "CRITICAL",
       "evidence": "Base64: L488: decoded_bytes = base64.b64decode(base64_encoded)\nSubprocess: L80: return subprocess.call(['which', name], | L100: p = subprocess.Popen(['pbcopy', 'w'], | L105: p = subprocess.Popen(['pbpaste', 'r'],"
     },
-    {
-      "package": "pytest",
-      "file": "_pytest/_py/path.py",
-      "check": "Downloads and executes remote code",
-      "severity": "CRITICAL",
-      "evidence": "L1153: exec(f.read(), mod.__dict__)"
-    },
-    {
-      "package": "pytest",
-      "file": "_pytest/capture.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L483: os.dup2(self.targetfd_invalid, targetfd) | L522: os.dup2(self.tmpfile.fileno(), self.targetfd) | L532: os.dup2(self.targetfd_save, self.targetfd)"
-    },
-    {
-      "package": "pytest",
-      "file": "_pytest/config/__init__.py",
-      "check": "Reverse shell / bind shell pattern",
-      "severity": "CRITICAL",
-      "evidence": "L260: os.dup2(devnull, sys.stdout.fileno())"
-    },
     {
       "package": "python-dateutil",
       "file": "dateutil/__init__.py",
@@ -527,6 +499,34 @@
       "severity": "CRITICAL",
       "evidence": "L13: __import__(__package__ + '.linalg') | L14: __import__(__package__ + '.fft')"
     },
+    {
+      "package": "scipy",
+      "file": "scipy/_external/array_api_compat/cupy/__init__.py",
+      "check": "Downloads and executes remote code",
+      "severity": "CRITICAL",
+      "evidence": "L12: __import__(__package__ + '.linalg') | L13: __import__(__package__ + '.fft')"
+    },
+    {
+      "package": "scipy",
+      "file": "scipy/_external/array_api_compat/dask/array/__init__.py",
+      "check": "Downloads and executes remote code",
+      "severity": "CRITICAL",
+      "evidence": "L16: __import__(__package__ + '.linalg') | L17: __import__(__package__ + '.fft')"
+    },
+    {
+      "package": "scipy",
+      "file": "scipy/_external/array_api_compat/numpy/__init__.py",
+      "check": "Downloads and executes remote code",
+      "severity": "CRITICAL",
+      "evidence": "L23: __import__(__package__ + \".linalg\") | L25: __import__(__package__ + \".fft\")"
+    },
+    {
+      "package": "scipy",
+      "file": "scipy/_external/array_api_compat/torch/__init__.py",
+      "check": "Downloads and executes remote code",
+      "severity": "CRITICAL",
+      "evidence": "L13: __import__(__package__ + '.linalg') | L14: __import__(__package__ + '.fft')"
+    },
     {
       "package": "sentencepiece",
       "file": "sentencepiece/__init__.py",
@@ -814,6 +814,13 @@
       "severity": "CRITICAL",
       "evidence": "L67: input_gguf=\"/tmp/in.gguf\","
     },
+    {
+      "package": "unsloth-zoo",
+      "file": "tests/test_mlx_save_export_regressions.py",
+      "check": "Writes to /tmp and executes (staged dropper)",
+      "severity": "CRITICAL",
+      "evidence": "L164: temporary_location=\"/tmp/ignored\","
+    },
     {
       "package": "unsloth-zoo",
       "file": "tests/test_upstream_pinned_symbols_transformers.py",
@@ -933,13 +940,6 @@
       "severity": "HIGH",
       "evidence": "Key: L187: \"-----BEGIN PUBLIC KEY-----\", | L188: \"-----BEGIN RSA PUBLIC KEY-----\",\nNetwork: L225: http_client: httpx.AsyncClient | None = None, | L411: else httpx.AsyncClient(timeout=httpx.Timeout(10.0))"
     },
-    {
-      "package": "hypothesis",
-      "file": "hypothesis/internal/scrutineer.py",
-      "check": "Anti-analysis/sandbox evasion + suspicious behavior",
-      "severity": "HIGH",
-      "evidence": "Anti: L76: return sys.gettrace() is None | L113: sys.settrace(self.trace) | L136: sys.settrace(None)"
-    },
     {
       "package": "ipython",
       "file": "IPython/core/debugger.py",
@@ -982,20 +982,6 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L709: return compile(source, filename, \"exec\")\nExec: L1228: exec(code, namespace)"
     },
-    {
-      "package": "kgb",
-      "file": "kgb/spies.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L934: eval(compile(func_code_str, '<string>', 'exec'),\nExec: L934: eval(compile(func_code_str, '<string>', 'exec'),"
-    },
-    {
-      "package": "langid",
-      "file": "langid/train/common.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L44: yield marshal.load(t)\nExec: L85: key = eval(row[0])"
-    },
     {
       "package": "matplotlib",
       "file": "matplotlib/sphinxext/plot_directive.py",
@@ -1108,20 +1094,6 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L45: mod = __import__(module_name, None, None, ['__all__'])\nExec: L154: exec(f.read(), custom_namespace)"
     },
-    {
-      "package": "pytest",
-      "file": "_pytest/_py/path.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L626: mod = __import__(hashtype) | L1118: __import__(modname)\nExec: L1153: exec(f.read(), mod.__dict__)"
-    },
-    {
-      "package": "pytest",
-      "file": "_pytest/assertion/rewrite.py",
-      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
-      "severity": "HIGH",
-      "evidence": "Obfusc: L393: co = marshal.load(fp) | L395: trace(f\"_read_pyc({source}): marshal.load error {e}\")\nExec: L188: exec(co, module.__dict__)"
-    },
     {
       "package": "scikit-learn",
       "file": "sklearn/externals/array_api_compat/torch/__init__.py",
@@ -1318,6 +1290,13 @@
       "severity": "HIGH",
       "evidence": "Obfusc: L3078: module = __import__('transformers', fromlist=[model_class_name])\nExec: L2960: exec(f\"from transformers.modeling_utils import ({', '.join(functions)})\", locals(), globals()) | L3006: exec(save_pretrained, globals(), functions)"
     },
+    {
+      "package": "unsloth-zoo",
+      "file": "tests/test_mlx_trainer_internals.py",
+      "check": "Advanced obfuscation (marshal/compile/zlib) + exec/eval",
+      "severity": "HIGH",
+      "evidence": "Obfusc: L430: assert ppl == pytest.approx(__import__(\"math\").exp(2.5))\nExec: L408: def eval(self):"
+    },
     {
       "package": "werkzeug",
       "file": "werkzeug/routing/rules.py",
diff --git a/scripts/verify_import_hoist.py b/scripts/verify_import_hoist.py
index b8cb0573fe..b4c908b0cb 100644
--- a/scripts/verify_import_hoist.py
+++ b/scripts/verify_import_hoist.py
@@ -581,9 +581,16 @@ def compare(before_src: str, after_src: str, path: str) -> list[tuple[str, str]]
             )
 
     # 3. TARGET-CHANGED (same scope+name resolves to a different import target)
+    #    Only a *swap* is dangerous: a BEFORE target that is no longer reachable in
+    #    AFTER means a reference was silently re-pointed. A pure superset growth
+    #    (tbefore <= tafter) is the benign `import pkg.subA` + `import pkg.subB`
+    #    case: both statements bind the same top-level name `pkg` to the same
+    #    package object and only *add* submodule attributes (e.g. adding
+    #    `import urllib.error` next to `import urllib.request`). Nothing the name
+    #    resolved to before is lost, so no reference is re-pointed -- skip it.
     for key, tafter in b["target_by_use"].items():
         tbefore = a["target_by_use"].get(key)
-        if tbefore and tbefore != tafter:
+        if tbefore and tbefore != tafter and (tbefore - tafter):
             findings.append(
                 (
                     "BLOCKER",
diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 2b9517692f..4dca4db768 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -26,6 +26,7 @@ from utils.hardware import (
 )
 from core.inference.audio_codecs import AudioCodecManager
 from core.inference.runtime_context import runtime_context_length
+from core.inference.message_content import content_to_text
 from io import StringIO
 import structlog
 from loggers import get_logger
@@ -1018,7 +1019,7 @@ class InferenceBackend:
         user_message = ""
         if messages and messages[-1]["role"] == "user":
             import re
-            user_message = messages[-1]["content"]
+            user_message = content_to_text(messages[-1]["content"])
             user_message = re.sub(r"<img[^>]*>", "", user_message).strip()
 
         if not user_message:
@@ -1181,7 +1182,7 @@ class InferenceBackend:
         if messages:
             for msg in reversed(messages):
                 if msg["role"] == "user" and msg.get("content"):
-                    user_text = msg["content"]
+                    user_text = content_to_text(msg["content"])
                     break
 
         # ASR-specific default system prompt if none set
@@ -1713,7 +1714,7 @@ class InferenceBackend:
 
         for msg in messages:
             role = msg.get("role", "")
-            content = msg.get("content", "")
+            content = content_to_text(msg.get("content", ""))
 
             if role in ["system", "user", "assistant"] and content.strip():
                 if role == last_role:
@@ -1801,7 +1802,7 @@ class InferenceBackend:
 
         for msg in messages:
             role = msg["role"]
-            content = msg["content"]
+            content = content_to_text(msg["content"])
             formatted += f"<|start_header_id|>{role}<|end_header_id|>\n\n{content}<|eot_id|>"
 
         formatted += "<|start_header_id|>assistant<|end_header_id|>\n\n"
@@ -1817,14 +1818,14 @@ class InferenceBackend:
 
         for msg in messages:
             if msg["role"] == "system":
-                system_msg = msg["content"]
+                system_msg = content_to_text(msg["content"])
             else:
                 conversation.append(msg)
 
         i = 0
         while i < len(conversation):
             if conversation[i]["role"] == "user":
-                user_content = conversation[i]["content"]
+                user_content = content_to_text(conversation[i]["content"])
 
                 if system_msg and i == 0:
                     user_content = f"{system_msg}\n\n{user_content}"
@@ -1832,7 +1833,7 @@ class InferenceBackend:
                 formatted += f"[INST] {user_content} [/INST]"
 
                 if i + 1 < len(conversation) and conversation[i + 1]["role"] == "assistant":
-                    formatted += f" {conversation[i + 1]['content']}</s>"
+                    formatted += f" {content_to_text(conversation[i + 1]['content'])}</s>"
                     i += 2
                 else:
                     formatted += " "
@@ -1848,7 +1849,7 @@ class InferenceBackend:
 
         for msg in messages:
             role = msg["role"]
-            content = msg["content"]
+            content = content_to_text(msg["content"])
             formatted += f"<|im_start|>{role}\n{content}<|im_end|>\n"
 
         formatted += "<|im_start|>assistant\n"
@@ -1860,16 +1861,17 @@ class InferenceBackend:
         system_msg = None
 
         for msg in messages:
+            content = content_to_text(msg["content"])
             if msg["role"] == "system":
-                system_msg = msg["content"]
+                system_msg = content
             elif msg["role"] == "user":
                 if system_msg:
-                    formatted += f"### Instruction:\n{system_msg}\n\n### Input:\n{msg['content']}\n\n### Response:\n"
+                    formatted += f"### Instruction:\n{system_msg}\n\n### Input:\n{content}\n\n### Response:\n"
                     system_msg = None
                 else:
-                    formatted += f"### Human:\n{msg['content']}\n\n### Assistant:\n"
+                    formatted += f"### Human:\n{content}\n\n### Assistant:\n"
             elif msg["role"] == "assistant":
-                formatted += f"{msg['content']}\n\n"
+                formatted += f"{content}\n\n"
 
         return formatted
 
@@ -1879,7 +1881,7 @@ class InferenceBackend:
 
         for msg in messages:
             role = msg["role"].title()
-            content = msg["content"]
+            content = content_to_text(msg["content"])
             formatted += f"{role}: {content}\n"
 
         formatted += "Assistant: "
diff --git a/studio/backend/core/inference/llama_cpp.py b/studio/backend/core/inference/llama_cpp.py
index eedda112f4..83b56c0091 100644
--- a/studio/backend/core/inference/llama_cpp.py
+++ b/studio/backend/core/inference/llama_cpp.py
@@ -7861,6 +7861,15 @@ class LlamaCppBackend:
         _accumulated_completion_tokens = 0
         _accumulated_predicted_ms = 0.0
         _accumulated_predicted_n = 0
+        # GGUF buffers reasoning; emit server-side timing before answer text.
+        _reasoning_started_at: Optional[float] = None
+        _reasoning_summary_emitted = False
+
+        def _reasoning_summary_event(started_at: float) -> dict:
+            return {
+                "type": "reasoning_summary",
+                "duration_ms": round((time.monotonic() - started_at) * 1000.0),
+            }
 
         def _strip_tool_markup(
             text: str,
@@ -7998,6 +8007,9 @@ class LlamaCppBackend:
                 content_buffer = ""  # Raw content held during BUFFERING
                 content_accum = ""  # All content tokens (for tool parsing)
                 reasoning_accum = ""
+                # Time each reasoning pass so final answers can replace tool timing.
+                _reasoning_started_at = None
+                _reasoning_summary_emitted = False
                 cumulative_display = ""  # Cumulative yielded text (with <think>)
                 in_thinking = False
                 has_content_tokens = False
@@ -8172,6 +8184,8 @@ class LlamaCppBackend:
                                 # between tool iterations).
                                 reasoning = delta.get("reasoning_content", "")
                                 if reasoning:
+                                    if _reasoning_started_at is None:
+                                        _reasoning_started_at = time.monotonic()
                                     reasoning_accum += reasoning
                                     if detect_state == _S_STREAMING:
                                         if not in_thinking:
@@ -8187,6 +8201,13 @@ class LlamaCppBackend:
                                 # ── Content tokens ──
                                 token = delta.get("content", "")
                                 if token:
+                                    # First answer token ends reasoning.
+                                    if (
+                                        _reasoning_started_at is not None
+                                        and not _reasoning_summary_emitted
+                                    ):
+                                        _reasoning_summary_emitted = True
+                                        yield _reasoning_summary_event(_reasoning_started_at)
                                     has_content_tokens = True
                                     content_accum += token
 
@@ -8284,9 +8305,10 @@ class LlamaCppBackend:
                                     ),
                                 }
                         elif reasoning_accum and not has_content_tokens:
-                            # Reasoning-only response: show reasoning as plain
-                            # text, matching the final streaming pass for
-                            # models that put everything in reasoning.
+                            # Reasoning-only reply: show it as plain text.
+                            if _reasoning_started_at is not None and not _reasoning_summary_emitted:
+                                _reasoning_summary_emitted = True
+                                yield _reasoning_summary_event(_reasoning_started_at)
                             cumulative_display = reasoning_accum
                             if not _suppress_visible_output:
                                 yield {
@@ -8695,6 +8717,8 @@ class LlamaCppBackend:
         in_thinking = False
         has_content_tokens = False
         reasoning_text = ""
+        _final_reasoning_started_at: Optional[float] = None
+        _final_reasoning_summary_emitted = False
         _metadata_usage = None
         _metadata_timings = None
         _metadata_finish_reason = None
@@ -8723,6 +8747,12 @@ class LlamaCppBackend:
                             continue
                         if line == "data: [DONE]":
                             if in_thinking:
+                                if (
+                                    _final_reasoning_started_at is not None
+                                    and not _final_reasoning_summary_emitted
+                                ):
+                                    _final_reasoning_summary_emitted = True
+                                    yield _reasoning_summary_event(_final_reasoning_started_at)
                                 if has_content_tokens:
                                     cumulative += "</think>"
                                     yield {
@@ -8755,6 +8785,8 @@ class LlamaCppBackend:
 
                                 reasoning = delta.get("reasoning_content", "")
                                 if reasoning:
+                                    if _final_reasoning_started_at is None:
+                                        _final_reasoning_started_at = time.monotonic()
                                     reasoning_text += reasoning
                                     if not in_thinking:
                                         cumulative += "<think>"
@@ -8764,6 +8796,12 @@ class LlamaCppBackend:
 
                                 token = delta.get("content", "")
                                 if token:
+                                    if (
+                                        _final_reasoning_started_at is not None
+                                        and not _final_reasoning_summary_emitted
+                                    ):
+                                        _final_reasoning_summary_emitted = True
+                                        yield _reasoning_summary_event(_final_reasoning_started_at)
                                     has_content_tokens = True
                                     if in_thinking:
                                         cumulative += "</think>"
diff --git a/studio/backend/core/inference/message_content.py b/studio/backend/core/inference/message_content.py
new file mode 100644
index 0000000000..b7c499a087
--- /dev/null
+++ b/studio/backend/core/inference/message_content.py
@@ -0,0 +1,38 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Normalize chat-message `content` (string or OpenAI multimodal list) to text.
+
+String-only formatting paths called string ops directly on `content` and broke
+on the list form (#4383). `content_to_text` collapses either shape to a string,
+dropping non-text parts. No heavy imports, so it is unit-testable alone.
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+
+def content_to_text(content: Any) -> str:
+    """Plain text of a `content`: str unchanged, list/tuple text parts newline-joined
+    (non-text dropped), None to "", else str(content)."""
+    if content is None:
+        return ""
+    if isinstance(content, str):
+        return content
+    if isinstance(content, (list, tuple)):
+        parts = []
+        for item in content:
+            if isinstance(item, str):
+                if item:
+                    parts.append(item)
+            elif isinstance(item, dict):
+                # Skip non-text parts (image_url, input_audio, ...).
+                part_type = item.get("type")
+                if part_type is not None and part_type != "text":
+                    continue
+                text = item.get("text")
+                if isinstance(text, str) and text:
+                    parts.append(text)
+        return "\n".join(parts)
+    return str(content)
diff --git a/studio/backend/core/inference/tool_call_parser.py b/studio/backend/core/inference/tool_call_parser.py
index 8d5d45269e..ca3d1e4cbc 100644
--- a/studio/backend/core/inference/tool_call_parser.py
+++ b/studio/backend/core/inference/tool_call_parser.py
@@ -7,25 +7,31 @@ Tolerates missing closing tags in either ``<tool_call>{json}</tool_call>``
 or ``<function=name><parameter=k>v...`` shape.
 """
 
-import json
-import re
+from core import tool_healing as _tool_healing
 
 
-# _TOOL_CLOSED_PATS: closed pairs only. _TOOL_ALL_PATS: also trailing unclosed
-# runs so truncated tails don't leak markup. The [\w-] name set matches OpenAI's
-# so hyphenated MCP tool names (mcp__srv__list-issues) parse like built-ins.
-_TOOL_CLOSED_PATS = [
-    re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
-    re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL),
-]
-_TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
-    re.compile(r"<tool_call>.*$", re.DOTALL),
-    re.compile(r"<function=[\w-]+>.*$", re.DOTALL),
-]
+_TOOL_ALL_PATS = _tool_healing._TOOL_ALL_PATS
+
+
+def parse_tool_calls_from_text(
+    content: str,
+    *,
+    id_offset: int = 0,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    return _tool_healing.parse_tool_calls_from_text(
+        content,
+        id_offset = id_offset,
+        allow_incomplete = allow_incomplete,
+    )
+
+
+def strip_tool_markup(text: str, *, final: bool = False) -> str:
+    return _tool_healing.strip_tool_call_markup(text, final = final)
 
 
 # Prefixes the streaming buffer watches for to gate in-progress text.
-TOOL_XML_SIGNALS = ("<tool_call>", "<function=")
+TOOL_XML_SIGNALS = ("<tool_call>", "<|tool_call>", "<function=")
 
 
 # Nudges + error prefixes shared by the GGUF and safetensors loops.
@@ -74,199 +80,6 @@ RAG_SEARCH_CAP_NUDGE = (
 )
 
 
-# Pre-compiled patterns reused by ``parse_tool_calls_from_text``.
-_TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
-_TC_FUNC_START_RE = re.compile(r"<function=([\w-]+)>\s*")
-_TC_END_TAG_RE = re.compile(r"</tool_call>")
-_TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
-# [\w-] so hyphenated MCP param names (issue-number) aren't dropped.
-_TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>\s*")
-_TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
-_PARAM_CLOSE_TAG = "</parameter>"
-_FUNC_CLOSE_TAG = "</function>"
-
-
-def _inside_open_parameter(content: str, pos: int) -> bool:
-    """Return True when ``pos`` falls inside an unclosed parameter value."""
-    last_param_start = -1
-    for match in _TC_PARAM_START_RE.finditer(content, 0, pos):
-        last_param_start = match.start()
-    if last_param_start < 0:
-        return False
-    last_param_close = content.rfind(_PARAM_CLOSE_TAG, 0, pos)
-    last_func_close = content.rfind(_FUNC_CLOSE_TAG, 0, pos)
-    return last_param_start > max(last_param_close, last_func_close)
-
-
-def strip_tool_markup(text: str, *, final: bool = False) -> str:
-    """Strip tool-call XML from streamed text.
-
-    ``final=False`` only removes closed pairs (used during streaming so
-    in-progress XML stays buffered). ``final=True`` also removes a
-    trailing unclosed run and trims the result.
-    """
-    pats = _TOOL_ALL_PATS if final else _TOOL_CLOSED_PATS
-    for pat in pats:
-        text = pat.sub("", text)
-    return text.strip() if final else text
-
-
-def parse_tool_calls_from_text(
-    content: str,
-    *,
-    id_offset: int = 0,
-    allow_incomplete: bool = True,
-) -> list[dict]:
-    """Parse OpenAI-format ``tool_calls`` from model text.
-
-    Returns a list of ``{"id", "type", "function": {"name", "arguments"}}``
-    dicts. ``arguments`` is always a JSON string so callers can hand it
-    straight back into an OpenAI-style response.
-
-    Handles two shapes:
-
-    - JSON inside ``<tool_call>`` tags:
-      ``<tool_call>{"name":"web_search","arguments":{"query":"..."}}</tool_call>``
-    - XML-style function blocks:
-      ``<function=name><parameter=k>v</parameter></function>``
-
-    ``allow_incomplete=True`` keeps the historical healing behavior for
-    missing closing tags. ``allow_incomplete=False`` accepts only
-    well-formed wrappers so disabled Auto-Heal can still parse valid
-    local tool protocol without repairing truncated output.
-    """
-    tool_calls: list[dict] = []
-
-    # Pattern 1: <tool_call>{json}. Balanced-brace scan, skipping braces in
-    # JSON strings.
-    for m in _TC_JSON_START_RE.finditer(content):
-        brace_start = m.end() - 1  # opening {
-        depth, i = 0, brace_start
-        in_string = False
-        while i < len(content):
-            ch = content[i]
-            if in_string:
-                if ch == "\\" and i + 1 < len(content):
-                    i += 2
-                    continue
-                if ch == '"':
-                    in_string = False
-            elif ch == '"':
-                in_string = True
-            elif ch == "{":
-                depth += 1
-            elif ch == "}":
-                depth -= 1
-                if depth == 0:
-                    break
-            i += 1
-        if depth != 0:
-            continue
-        if not allow_incomplete:
-            tail_after_json = content[i + 1 :].lstrip()
-            if _TC_END_TAG_RE.match(tail_after_json) is None:
-                continue
-        json_str = content[brace_start : i + 1]
-        try:
-            obj = json.loads(json_str)
-            tc = {
-                "id": f"call_{id_offset + len(tool_calls)}",
-                "type": "function",
-                "function": {
-                    "name": obj.get("name", ""),
-                    "arguments": obj.get("arguments", {}),
-                },
-            }
-            if isinstance(tc["function"]["arguments"], dict):
-                tc["function"]["arguments"] = json.dumps(tc["function"]["arguments"])
-            tool_calls.append(tc)
-        except (json.JSONDecodeError, ValueError):
-            pass
-
-    # Pattern 2: <function=name><parameter=k>v... -- closing tags optional;
-    # </function> isn't a body boundary since code values can contain it.
-    if not tool_calls:
-        func_starts = [
-            fm
-            for fm in _TC_FUNC_START_RE.finditer(content)
-            if not _inside_open_parameter(content, fm.start())
-        ]
-        for idx, fm in enumerate(func_starts):
-            func_name = fm.group(1)
-            body_start = fm.end()
-            next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
-            end_tag = _TC_END_TAG_RE.search(content[body_start:])
-            if end_tag:
-                body_end = body_start + end_tag.start()
-            else:
-                body_end = len(content)
-            body_end = min(body_end, next_func)
-            body = content[body_start:body_end]
-            if not allow_incomplete:
-                # Bound the body at the closing </function> tag rather than
-                # the end of the response, so a complete call followed by
-                # trailing prose is still accepted (matching the JSON-style
-                # <tool_call> path, which already tolerates trailing text).
-                # rfind picks the last </function>, so a literal </function>
-                # inside a code parameter value stays in the body.
-                close_idx = body.rfind(_FUNC_CLOSE_TAG)
-                if close_idx < 0:
-                    continue
-                body = body[:close_idx]
-            else:
-                body = _TC_FUNC_CLOSE_RE.sub("", body)
-
-            arguments: dict = {}
-            param_starts = list(_TC_PARAM_START_RE.finditer(body))
-            if len(param_starts) == 1:
-                # Single param: take everything to body end so an embedded
-                # </parameter> in code strings is preserved.
-                pm = param_starts[0]
-                val = body[pm.end() :]
-                if not allow_incomplete:
-                    stripped_val = val.rstrip()
-                    if not stripped_val.endswith(_PARAM_CLOSE_TAG):
-                        continue
-                    val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
-                else:
-                    val = _TC_PARAM_CLOSE_RE.sub("", val)
-                arguments[pm.group(1)] = val.strip()
-            else:
-                valid_params = True
-                for pidx, pm in enumerate(param_starts):
-                    param_name = pm.group(1)
-                    val_start = pm.end()
-                    next_param = (
-                        param_starts[pidx + 1].start()
-                        if pidx + 1 < len(param_starts)
-                        else len(body)
-                    )
-                    val = body[val_start:next_param]
-                    if not allow_incomplete:
-                        stripped_val = val.rstrip()
-                        if not stripped_val.endswith(_PARAM_CLOSE_TAG):
-                            valid_params = False
-                            break
-                        val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
-                    else:
-                        val = _TC_PARAM_CLOSE_RE.sub("", val)
-                    arguments[param_name] = val.strip()
-                if not valid_params:
-                    continue
-
-            tc = {
-                "id": f"call_{id_offset + len(tool_calls)}",
-                "type": "function",
-                "function": {
-                    "name": func_name,
-                    "arguments": json.dumps(arguments),
-                },
-            }
-            tool_calls.append(tc)
-
-    return tool_calls
-
-
 def has_tool_signal(text: str) -> bool:
     """Return True if ``text`` contains any tool-call XML signal."""
     return any(s in text for s in TOOL_XML_SIGNALS)
diff --git a/studio/backend/core/inference/tools.py b/studio/backend/core/inference/tools.py
index 6960310018..a5c193ff39 100644
--- a/studio/backend/core/inference/tools.py
+++ b/studio/backend/core/inference/tools.py
@@ -2545,14 +2545,24 @@ def _python_exec(
                         pass
     try:
         fd, tmp_path = tempfile.mkstemp(suffix = ".py", prefix = "studio_exec_", dir = workdir)
-        with os.fdopen(fd, "w") as f:
+        # utf-8 so non-ASCII in model-written code survives the OS default codec
+        # (Windows cp1252 would otherwise raise UnicodeEncodeError).
+        with os.fdopen(fd, "w", encoding = "utf-8") as f:
             f.write(code)
 
         safe_env = _build_bypass_env(workdir) if disable_sandbox else _build_safe_env(workdir)
+        if disable_sandbox:
+            # Match the sandboxed Python path without changing bypass shell I/O.
+            safe_env = dict(safe_env)
+            safe_env["PYTHONIOENCODING"] = "utf-8"
         popen_kwargs = dict(
             stdout = subprocess.PIPE,
             stderr = subprocess.STDOUT,
             text = True,
+            # Decode child output as utf-8 (it emits utf-8 via PYTHONIOENCODING);
+            # replace so non-ASCII output never crashes the read on Windows.
+            encoding = "utf-8",
+            errors = "replace",
             cwd = workdir,
             env = safe_env,
         )
diff --git a/studio/backend/core/rag/embeddings.py b/studio/backend/core/rag/embeddings.py
index 4e76e4fcaa..4c8d690302 100644
--- a/studio/backend/core/rag/embeddings.py
+++ b/studio/backend/core/rag/embeddings.py
@@ -46,6 +46,23 @@ def _device() -> str:
     return _TORCH_DEVICE.get(get_device(), "cpu")
 
 
+_torchao_stub_done = False
+
+
+def _install_torchao_stub_once() -> None:
+    """Neutralize torchao before importing sentence-transformers. On Windows ROCm,
+    torchao (pulled in by transformers.quantizers) imports an absent c10d backend
+    and aborts, dropping the embedder to llama-server. Workers stub it too; the
+    embedder runs in the main process. No-op elsewhere; runs once under ``_lock``."""
+    global _torchao_stub_done
+    if _torchao_stub_done:
+        return
+    _torchao_stub_done = True
+    from core._torchao_stub import install_torchao_windows_rocm_stub
+
+    install_torchao_windows_rocm_stub()
+
+
 def _get(model_name: str | None = None):
     """Cached SentenceTransformer, (re)loading on a name change. Loaded in fp16
     for a ~1.5x speedup at negligible accuracy loss."""
@@ -53,6 +70,7 @@ def _get(model_name: str | None = None):
     name = model_name or config.EMBEDDING_MODEL
     with _lock:
         if _model is None or _name != name:
+            _install_torchao_stub_once()
             from sentence_transformers import SentenceTransformer
 
             device = _device()
diff --git a/studio/backend/core/tool_healing.py b/studio/backend/core/tool_healing.py
index 973520d5cd..fe26d48c7f 100644
--- a/studio/backend/core/tool_healing.py
+++ b/studio/backend/core/tool_healing.py
@@ -1,14 +1,10 @@
 # SPDX-License-Identifier: AGPL-3.0-only
 # Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
-"""Tool-call XML parsing and stripping helpers.
+"""Lightweight tool-call XML parsing and stripping helpers.
 
-Extracted verbatim from studio/backend/core/inference/llama_cpp.py so external
-inference servers can reuse the logic without importing the inference
+External inference servers import this module without pulling in the inference
 orchestrator, structlog, httpx, or the rest of the studio backend.
-
-Regexes and bodies are byte-for-byte identical to the original; any change must
-preserve that. test_tool_healing_extraction_is_exact.py verifies via AST.
 """
 
 import json
@@ -19,86 +15,370 @@ import re
 # issue-number) parse alongside the built-ins.
 _TOOL_CLOSED_PATS = [
     re.compile(r"<tool_call>.*?</tool_call>", re.DOTALL),
+    re.compile(r"<\|tool_call>.*?<tool_call\|>", re.DOTALL),
+    re.compile(r"<tool_call\|>"),
     re.compile(r"<function=[\w-]+>.*?</function>", re.DOTALL),
 ]
 _TOOL_ALL_PATS = _TOOL_CLOSED_PATS + [
     re.compile(r"<tool_call>.*$", re.DOTALL),
+    re.compile(r"<\|tool_call>.*$", re.DOTALL),
     re.compile(r"<function=[\w-]+>.*$", re.DOTALL),
 ]
 
 # Pre-compiled patterns for tool-call XML parsing.
 _TC_JSON_START_RE = re.compile(r"<tool_call>\s*\{")
+_TC_GEMMA_START_RE = re.compile(r"<\|tool_call>call:([\w-]+)\s*\{")
 _TC_FUNC_START_RE = re.compile(r"<function=([\w-]+)>\s*")
 _TC_END_TAG_RE = re.compile(r"</tool_call>")
+_TC_GEMMA_END_TAG_RE = re.compile(r"<tool_call\|>")
 _TC_FUNC_CLOSE_RE = re.compile(r"\s*</function>\s*$")
 _TC_PARAM_START_RE = re.compile(r"<parameter=([\w-]+)>\s*")
 _TC_PARAM_CLOSE_RE = re.compile(r"\s*</parameter>\s*$")
+_GEMMA_QUOTE = '<|"|>'
+_PARAM_CLOSE_TAG = "</parameter>"
+_FUNC_CLOSE_TAG = "</function>"
+# A bare (unquoted) Gemma value ends at `}` or at a comma that begins the next
+# `key:` pair. A comma NOT followed by a key token is part of the value (e.g.
+# `location:New York, NY`), so it must not terminate the value. The key token
+# must be identifier-shaped (start with a letter or underscore); a comma
+# followed by digits-then-colon is value text such as a timestamp or ratio
+# (`meet at 10:00, 11:00 tomorrow`), not a new key.
+_GEMMA_NEXT_KEY_RE = re.compile(r"\s*[A-Za-z_][\w-]*\s*:")
 
 
-def parse_tool_calls_from_text(content: str) -> list[dict]:
-    """
-    Parse tool calls from XML markup in content text.
+def _balanced_brace_end(
+    content: str,
+    brace_start: int,
+    *,
+    gemma_quotes: bool = False,
+) -> int:
+    depth = 0
+    i = brace_start
+    in_string = False
+    in_gemma_string = False
+    while i < len(content):
+        if gemma_quotes and not in_string and content.startswith(_GEMMA_QUOTE, i):
+            in_gemma_string = not in_gemma_string
+            i += len(_GEMMA_QUOTE)
+            continue
+        ch = content[i]
+        if in_gemma_string:
+            i += 1
+            continue
+        if in_string:
+            if ch == "\\" and i + 1 < len(content):
+                i += 2
+                continue
+            if ch == '"':
+                in_string = False
+        elif ch == '"':
+            in_string = True
+        elif ch == "{":
+            depth += 1
+        elif ch == "}":
+            depth -= 1
+            if depth == 0:
+                return i
+        i += 1
+    return -1
+
+
+def _balanced_bracket_end(src: str, start: int) -> int:
+    """Index of the ``]`` matching the ``[`` at ``start``, or -1. Tracks nested
+    ``[]``/``{}`` and double-quoted strings."""
+    depth = 0
+    i = start
+    in_string = False
+    while i < len(src):
+        ch = src[i]
+        if in_string:
+            if ch == "\\" and i + 1 < len(src):
+                i += 2
+                continue
+            if ch == '"':
+                in_string = False
+        elif ch == '"':
+            in_string = True
+        elif ch in "[{":
+            depth += 1
+        elif ch in "]}":
+            depth -= 1
+            if depth == 0:
+                return i
+        i += 1
+    return -1
+
+
+def _split_top_level_commas(src: str) -> list:
+    """Split on commas that are not inside a nested ``[]``/``{}`` or a string."""
+    parts: list[str] = []
+    depth = 0
+    in_string = False
+    start = 0
+    i = 0
+    while i < len(src):
+        ch = src[i]
+        if in_string:
+            if ch == "\\" and i + 1 < len(src):
+                i += 2
+                continue
+            if ch == '"':
+                in_string = False
+        elif ch == '"':
+            in_string = True
+        elif ch in "[{":
+            depth += 1
+        elif ch in "]}":
+            depth -= 1
+        elif ch == "," and depth == 0:
+            parts.append(src[start:i])
+            start = i + 1
+        i += 1
+    parts.append(src[start:])
+    return parts
+
+
+def _quote_gemma_array_elements(body: str) -> str:
+    """Normalise the elements of a Gemma array value so json.loads succeeds.
+
+    Gemma may emit ``labels:[bug,ui]`` without per-element quotes, or arrays of
+    objects (``items:[{path:a}]``) whose keys/values also lack quotes; left
+    as-is json.loads fails and the whole call is dropped. Bare string elements
+    are quoted, object and nested-array elements are normalised recursively, and
+    quoted strings (already normalised from ``<|"|>``), numbers, and JSON
+    literals are preserved."""
+    out: list[str] = []
+    for element in _split_top_level_commas(body):
+        stripped = element.strip()
+        if not stripped or stripped[0] == '"':
+            out.append(element)
+            continue
+        if stripped[0] == "{":
+            # Object element: quote its keys/bare values like a top-level object.
+            out.append(_quote_gemma_object_keys(stripped))
+            continue
+        if stripped[0] == "[":
+            # Nested array: normalise its elements too.
+            inner_end = _balanced_bracket_end(stripped, 0)
+            if inner_end == len(stripped) - 1:
+                out.append("[" + _quote_gemma_array_elements(stripped[1:inner_end]) + "]")
+            else:
+                out.append(element)
+            continue
+        try:
+            json.loads(stripped)
+            out.append(element)
+        except (json.JSONDecodeError, ValueError):
+            out.append(json.dumps(stripped))
+    return ",".join(out)
+
+
+def _normalise_gemma_quoted_strings(src: str) -> str:
+    parts: list[str] = []
+    i = 0
+    while i < len(src):
+        if not src.startswith(_GEMMA_QUOTE, i):
+            parts.append(src[i])
+            i += 1
+            continue
+        end = src.find(_GEMMA_QUOTE, i + len(_GEMMA_QUOTE))
+        if end < 0:
+            parts.append(src[i:])
+            break
+        raw_value = src[i + len(_GEMMA_QUOTE) : end]
+        parts.append(json.dumps(raw_value))
+        i = end + len(_GEMMA_QUOTE)
+    return "".join(parts)
+
+
+def _quote_gemma_object_keys(src: str) -> str:
+    parts: list[str] = []
+    i = 0
+    in_string = False
+    while i < len(src):
+        ch = src[i]
+        if in_string:
+            parts.append(ch)
+            if ch == "\\" and i + 1 < len(src):
+                parts.append(src[i + 1])
+                i += 2
+                continue
+            if ch == '"':
+                in_string = False
+            i += 1
+            continue
+        if ch == '"':
+            in_string = True
+            parts.append(ch)
+            i += 1
+            continue
+        if ch not in "{,":
+            parts.append(ch)
+            i += 1
+            continue
+
+        parts.append(ch)
+        i += 1
+        key_start = i
+        while i < len(src) and src[i].isspace():
+            i += 1
+        key_name_start = i
+        while i < len(src) and (src[i].isalnum() or src[i] in "_-"):
+            i += 1
+        key_name = src[key_name_start:i]
+        colon_pos = i
+        while colon_pos < len(src) and src[colon_pos].isspace():
+            colon_pos += 1
+        if key_name and colon_pos < len(src) and src[colon_pos] == ":":
+            parts.append(src[key_start:key_name_start])
+            parts.append(json.dumps(key_name))
+            parts.append(src[i:colon_pos])
+            parts.append(":")
+            i = colon_pos + 1
+            # Gemma may emit bare string values ({unit:celsius}); quote them so
+            # json.loads succeeds. JSON scalars/objects/arrays/quoted stay as-is.
+            ws = i
+            while i < len(src) and src[i].isspace():
+                i += 1
+            parts.append(src[ws:i])
+            if i < len(src) and src[i] == "[":
+                # Array value: quote bare string elements (e.g. labels:[bug,ui])
+                # so json.loads succeeds instead of dropping the call.
+                arr_end = _balanced_bracket_end(src, i)
+                if arr_end < 0:
+                    parts.append(src[i:])
+                    i = len(src)
+                else:
+                    parts.append("[" + _quote_gemma_array_elements(src[i + 1 : arr_end]) + "]")
+                    i = arr_end + 1
+            elif i < len(src) and src[i] not in '"{':
+                v_start = i
+                # Consume the bare value up to `}` or a comma that starts the
+                # next key:value pair; a comma inside the value (e.g.
+                # `New York, NY`) does not terminate it.
+                while i < len(src):
+                    if src[i] == "}":
+                        break
+                    if src[i] == "," and _GEMMA_NEXT_KEY_RE.match(src, i + 1):
+                        break
+                    i += 1
+                raw = src[v_start:i]
+                try:
+                    json.loads(raw.strip())
+                    parts.append(raw)
+                except (json.JSONDecodeError, ValueError):
+                    parts.append(json.dumps(raw.strip()) if raw.strip() else raw)
+        else:
+            parts.append(src[key_start:i])
+    return "".join(parts)
+
+
+def _gemma_arguments_to_json(args_src: str) -> dict:
+    """Parse Gemma 4's native call:name{key:value} argument object."""
+    args_src = args_src.strip()
+    if not args_src:
+        return {}
+    src = _normalise_gemma_quoted_strings(args_src)
+    src = "{" + src + "}"
+    src = _quote_gemma_object_keys(src)
+    return json.loads(src)
+
+
+def _inside_open_parameter(content: str, pos: int) -> bool:
+    """Return True when ``pos`` falls inside an unclosed parameter value."""
+    last_param_start = -1
+    for match in _TC_PARAM_START_RE.finditer(content, 0, pos):
+        last_param_start = match.start()
+    if last_param_start < 0:
+        return False
+    last_param_close = content.rfind(_PARAM_CLOSE_TAG, 0, pos)
+    last_func_close = content.rfind(_FUNC_CLOSE_TAG, 0, pos)
+    return last_param_start > max(last_param_close, last_func_close)
+
+
+def parse_tool_calls_from_text(
+    content: str,
+    *,
+    id_offset: int = 0,
+    allow_incomplete: bool = True,
+) -> list[dict]:
+    """Parse OpenAI-format tool calls from model text.
 
     Handles formats like:
       <tool_call>{"name":"web_search","arguments":{"query":"..."}}</tool_call>
+      <|tool_call>call:web_search{query:"..."}<tool_call|>
       <tool_call><function=web_search><parameter=query>...</parameter></function></tool_call>
-    Closing tags (</tool_call>, </function>, </parameter>) are all
-    optional since models frequently omit them.
     """
-    tool_calls = []
-
-    # Pattern 1: JSON inside <tool_call> tags. Balanced-brace extraction that
-    # skips braces inside JSON strings.
+    tool_calls: list[dict] = []
+    # Collect JSON- and Gemma-format candidates with their byte spans, then
+    # accept them in document order. Both order and spans matter:
+    #   * tools execute in returned order, so a call appearing earlier in the
+    #     text must be emitted first even across the two formats;
+    #   * a tool-call marker INSIDE another call's argument string is data, not a
+    #     call, so a candidate starting within an already accepted span is
+    #     skipped (covers a JSON marker nested in a Gemma arg and a Gemma marker
+    #     nested in a JSON arg alike, regardless of which format is outer).
+    candidates = []  # (start, brace_end, kind, match)
     for m in _TC_JSON_START_RE.finditer(content):
-        brace_start = m.end() - 1  # position of the opening {
-        depth, i = 0, brace_start
-        in_string = False
-        while i < len(content):
-            ch = content[i]
-            if in_string:
-                if ch == "\\" and i + 1 < len(content):
-                    i += 2  # skip escaped character
-                    continue
-                if ch == '"':
-                    in_string = False
-            elif ch == '"':
-                in_string = True
-            elif ch == "{":
-                depth += 1
-            elif ch == "}":
-                depth -= 1
-                if depth == 0:
-                    break
-            i += 1
-        if depth == 0:
-            json_str = content[brace_start : i + 1]
-            try:
-                obj = json.loads(json_str)
-                tc = {
-                    "id": f"call_{len(tool_calls)}",
-                    "type": "function",
-                    "function": {
-                        "name": obj.get("name", ""),
-                        "arguments": obj.get("arguments", {}),
-                    },
-                }
-                if isinstance(tc["function"]["arguments"], dict):
-                    tc["function"]["arguments"] = json.dumps(tc["function"]["arguments"])
-                tool_calls.append(tc)
-            except (json.JSONDecodeError, ValueError):
-                pass
+        # A marker that begins inside an open <function=...><parameter=...> value
+        # is that parameter's data, not its own call; skip it (same guard the
+        # XML-style parser below applies to nested <function= markers).
+        if _inside_open_parameter(content, m.start()):
+            continue
+        end = _balanced_brace_end(content, m.end() - 1)
+        if end >= 0:
+            candidates.append((m.start(), end, "json", m))
+    for m in _TC_GEMMA_START_RE.finditer(content):
+        if _inside_open_parameter(content, m.start()):
+            continue
+        end = _balanced_brace_end(content, m.end() - 1, gemma_quotes = True)
+        if end >= 0:
+            candidates.append((m.start(), end, "gemma", m))
+    candidates.sort(key = lambda c: c[0])
+
+    spans = [(s, e) for s, e, _kind, _m in candidates]
+    for idx, (start, end, kind, m) in enumerate(candidates):
+        # Skip a candidate nested inside another candidate's brace span: it is
+        # the enclosing call's argument data, not its own call. Checked against
+        # every candidate span (not only the ones that parsed successfully), so a
+        # marker inside an outer call that later fails to normalize is still
+        # never promoted to its own executable tool call.
+        if any(s <= start and end <= e for j, (s, e) in enumerate(spans) if j != idx):
+            continue
+        if not allow_incomplete:
+            tail = content[end + 1 :].lstrip()
+            close_re = _TC_END_TAG_RE if kind == "json" else _TC_GEMMA_END_TAG_RE
+            if close_re.match(tail) is None:
+                continue
+        try:
+            if kind == "json":
+                obj = json.loads(content[m.end() - 1 : end + 1])
+                name = obj.get("name", "")
+                arguments = obj.get("arguments", {})
+                if isinstance(arguments, dict):
+                    arguments = json.dumps(arguments)
+            else:
+                name = m.group(1)
+                arguments = json.dumps(_gemma_arguments_to_json(content[m.end() : end]))
+        except (json.JSONDecodeError, ValueError):
+            continue
+        tool_calls.append(
+            {
+                "id": f"call_{id_offset + len(tool_calls)}",
+                "type": "function",
+                "function": {"name": name, "arguments": arguments},
+            }
+        )
 
-    # Pattern 2: XML-style <function=name><parameter=key>value</parameter></function>
-    # All closing tags optional; models frequently omit them.
     if not tool_calls:
-        # Step 1: Find <function=name> positions and extract bodies. Use only
-        # </tool_call> or the next <function= as hard boundaries (</function>
-        # can appear in code values); trim a trailing </function> afterwards.
-        func_starts = list(_TC_FUNC_START_RE.finditer(content))
+        func_starts = [
+            fm
+            for fm in _TC_FUNC_START_RE.finditer(content)
+            if not _inside_open_parameter(content, fm.start())
+        ]
         for idx, fm in enumerate(func_starts):
             func_name = fm.group(1)
             body_start = fm.end()
-            # Boundaries: next <function= tag or </tool_call>
             next_func = func_starts[idx + 1].start() if idx + 1 < len(func_starts) else len(content)
             end_tag = _TC_END_TAG_RE.search(content[body_start:])
             if end_tag:
@@ -107,36 +387,52 @@ def parse_tool_calls_from_text(content: str) -> list[dict]:
                 body_end = len(content)
             body_end = min(body_end, next_func)
             body = content[body_start:body_end]
-            body = _TC_FUNC_CLOSE_RE.sub("", body)  # trim closing </function>
+            if not allow_incomplete:
+                close_idx = body.rfind(_FUNC_CLOSE_TAG)
+                if close_idx < 0:
+                    continue
+                body = body[:close_idx]
+            else:
+                body = _TC_FUNC_CLOSE_RE.sub("", body)
 
-            # Step 2: Extract parameters from body. For single-parameter
-            # functions, use body end as the only boundary to avoid matching
-            # </parameter> inside code strings.
-            arguments = {}
+            arguments: dict = {}
             param_starts = list(_TC_PARAM_START_RE.finditer(body))
             if len(param_starts) == 1:
-                # Value is everything after the tag to end of body, less a
-                # trailing </parameter>.
                 pm = param_starts[0]
                 val = body[pm.end() :]
-                val = _TC_PARAM_CLOSE_RE.sub("", val)
+                if not allow_incomplete:
+                    stripped_val = val.rstrip()
+                    if not stripped_val.endswith(_PARAM_CLOSE_TAG):
+                        continue
+                    val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
+                else:
+                    val = _TC_PARAM_CLOSE_RE.sub("", val)
                 arguments[pm.group(1)] = val.strip()
             else:
+                valid_params = True
                 for pidx, pm in enumerate(param_starts):
                     param_name = pm.group(1)
                     val_start = pm.end()
-                    # Value ends at next <parameter= or end of body
                     next_param = (
                         param_starts[pidx + 1].start()
                         if pidx + 1 < len(param_starts)
                         else len(body)
                     )
                     val = body[val_start:next_param]
-                    val = _TC_PARAM_CLOSE_RE.sub("", val)  # trim trailing </parameter>
+                    if not allow_incomplete:
+                        stripped_val = val.rstrip()
+                        if not stripped_val.endswith(_PARAM_CLOSE_TAG):
+                            valid_params = False
+                            break
+                        val = stripped_val[: -len(_PARAM_CLOSE_TAG)]
+                    else:
+                        val = _TC_PARAM_CLOSE_RE.sub("", val)
                     arguments[param_name] = val.strip()
+                if not valid_params:
+                    continue
 
             tc = {
-                "id": f"call_{len(tool_calls)}",
+                "id": f"call_{id_offset + len(tool_calls)}",
                 "type": "function",
                 "function": {
                     "name": func_name,
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 34bb4331a7..290b4090ec 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -24,9 +24,10 @@ from datetime import datetime, timezone
 from loggers import get_logger
 from dataclasses import dataclass, field
 from pathlib import Path
-from typing import Optional, Tuple, Any
+from typing import Optional, Tuple, Any, TYPE_CHECKING
 
-import matplotlib.pyplot as plt
+if TYPE_CHECKING:
+    import matplotlib.pyplot as plt
 from utils.hardware import prepare_gpu_selection
 from utils.native_path_leases import (
     native_path_secret_removed_for_child_start,
@@ -36,6 +37,30 @@ from utils.paths import outputs_root
 
 logger = get_logger(__name__)
 
+_pyplot = None
+_pyplot_failed = False
+
+
+def _load_pyplot():
+    """Lazily import matplotlib.pyplot (headless Agg); return it, or None if
+    matplotlib is unavailable. Deferred so a blocked native wheel (e.g. Windows
+    Smart App Control) never breaks server startup, only loss plotting.
+    """
+    global _pyplot, _pyplot_failed
+    if _pyplot is not None or _pyplot_failed:
+        return _pyplot
+    try:
+        import matplotlib
+
+        matplotlib.use("Agg")  # headless backend
+        import matplotlib.pyplot as plt
+
+        _pyplot = plt
+    except Exception as e:
+        _pyplot_failed = True
+        logger.warning("matplotlib unavailable; loss plots disabled", error = str(e))
+    return _pyplot
+
 
 def _coerce_seed(value, default = 3407) -> int:
     """Normalize None / non-int to `default` (transformers.set_seed(None) raises)."""
@@ -655,7 +680,7 @@ class TrainingBackend:
         plot = self._create_loss_plot(progress, theme)
         return (plot, progress)
 
-    def refresh_plot_for_theme(self, theme: str) -> Optional[plt.Figure]:
+    def refresh_plot_for_theme(self, theme: str) -> "Optional[plt.Figure]":
         """Refresh plot with new theme."""
         if theme and isinstance(theme, str) and theme in ["light", "dark"]:
             self.current_theme = theme
@@ -1090,8 +1115,14 @@ class TrainingBackend:
         self,
         progress: TrainingProgress,
         theme: str = "light",
-    ) -> plt.Figure:
-        """Create training loss plot with theme-aware styling."""
+    ) -> "Optional[plt.Figure]":
+        """Create training loss plot with theme-aware styling.
+
+        matplotlib is loaded lazily; returns None if it is unavailable.
+        """
+        plt = _load_pyplot()
+        if plt is None:
+            return None
         plt.close("all")
 
         LIGHT_STYLE = {
diff --git a/studio/backend/main.py b/studio/backend/main.py
index 6da6cf1a7d..d9bc74f883 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -917,24 +917,16 @@ async def _recipes_redirect(rest: str = ""):
     return _RedirectResponse(url = target, status_code = 308)
 
 
-_api_only = os.environ.get("UNSLOTH_API_ONLY") == "1"
-_cors_origins = ["*"]
-if _api_only:
-    _cors_origins = [
-        "tauri://localhost",  # Linux/macOS Tauri webview
-        "http://tauri.localhost",  # Windows Tauri webview
-        "http://localhost",  # dev fallback
-        "http://localhost:5173",  # Tauri dev/Vite
-        "http://127.0.0.1:5173",  # Tauri dev/Vite fallback
-    ]
-    _cors_origin_regex = None
-else:
-    _cors_origin_regex = None
+from utils.host_policy import cors_origins_for_mode  # noqa: E402
+
+_cors_origins = cors_origins_for_mode(
+    api_only = os.environ.get("UNSLOTH_API_ONLY") == "1",
+    secure = os.environ.get("UNSLOTH_SECURE") == "1",
+)
 
 app.add_middleware(
     CORSMiddleware,
     allow_origins = _cors_origins,
-    allow_origin_regex = _cors_origin_regex,
     allow_credentials = True,
     allow_methods = ["*"],
     allow_headers = ["*"],
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index b8432f588c..26825a472e 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -1102,6 +1102,8 @@ class ChoiceDelta(BaseModel):
 
     role: Optional[str] = None
     content: Optional[str] = None
+    reasoning_content: Optional[str] = None
+    tool_calls: Optional[list[dict]] = None
 
 
 OpenAIFinishReason = Literal["stop", "length", "tool_calls", "content_filter", "function_call"]
@@ -1137,6 +1139,8 @@ class CompletionMessage(BaseModel):
     role: Literal["assistant"] = "assistant"
     content: str
     refusal: Optional[str] = None
+    reasoning_content: Optional[str] = None
+    tool_calls: Optional[list[dict]] = None
 
 
 class CompletionChoice(BaseModel):
diff --git a/studio/backend/requirements/extras-no-deps.txt b/studio/backend/requirements/extras-no-deps.txt
index 23c61baa44..5830a47789 100644
--- a/studio/backend/requirements/extras-no-deps.txt
+++ b/studio/backend/requirements/extras-no-deps.txt
@@ -11,10 +11,12 @@ peft==0.18.1
 
 # TRL and related packages
 trl==0.23.1
-git+https://github.com/meta-pytorch/OpenEnv.git
 # executorch>=1.0.1               # 41.5 MB - no imports in unsloth/zoo/studio
 torch-c-dlpack-ext
 sentence_transformers==5.2.0
 transformers==4.57.6
 pytorch_tokenizers
 kernels==0.12.1
+# kernels<3.11 imports tomli as its tomllib fallback; --no-deps skips its own
+# marker dep, so list it here (no-op on the 3.12/3.13 default installs).
+tomli; python_version < "3.11"
diff --git a/studio/backend/requirements/extras.txt b/studio/backend/requirements/extras.txt
index 40737b0876..1baf2b6f2d 100644
--- a/studio/backend/requirements/extras.txt
+++ b/studio/backend/requirements/extras.txt
@@ -1,27 +1,11 @@
-# OpenEnv dependencies
-tomli
-tomli-w
-
-# ExecuTorch dependencies
-ruamel.yaml
-# coremltools                    # 10.2 MB - Apple CoreML, no imports in unsloth/zoo/studio
-expecttest
+# transitive dep of onnxruntime (via data-designer's pymupdf4llm)
 flatbuffers
-hydra-core
-hypothesis
-kgb
-parameterized
-pytest>=9.0.3,<10
-pytest-json-report
-pytest-rerunfailures>=16.2,<17
-pytest-xdist
-# Also needed by sentence_transformers (installed with --no-deps in extras-no-deps.txt)
+# Also needed by sentence_transformers (installed with --no-deps in extras-no-deps.txt);
+# librosa pulls it in too, but is skipped in no-torch mode.
 scikit-learn==1.7.1
 
 # Additional extras
-pybind11
-langid
-jiwer
+jiwer                            # WER/CER metrics for vision OCR save-merge benchmarks
 omegaconf
 einx
 pyloudnorm
@@ -39,17 +23,12 @@ ftfy
 importlib-resources
 librosa
 markdown2
-matplotlib
+matplotlib==3.10.9
 pystoi
 soundfile
 tensorboard
 torch-stoi
-evaluate
 timm
-transformers-cfg
-open_spiel
-addict
-easydict
 einops
 tabulate
 openai>=2.7.2
diff --git a/studio/backend/requirements/no-torch-runtime.txt b/studio/backend/requirements/no-torch-runtime.txt
index b0157cfea0..9796fc3a50 100644
--- a/studio/backend/requirements/no-torch-runtime.txt
+++ b/studio/backend/requirements/no-torch-runtime.txt
@@ -56,7 +56,7 @@ httpx
 httpcore
 certifi
 idna
-anyio>=3.0,<4.14.0  # 4.14+ breaks cancel scope on Py3.13 (#6483)
+anyio>=3.0,<4.14.0  # 4.14 asyncio cancel-scope RuntimeError on Py3.13 streaming (#6483); 4.13 unaffected
 sniffio
 h11
 
diff --git a/studio/backend/requirements/single-env/constraints.txt b/studio/backend/requirements/single-env/constraints.txt
index aad4c38664..0ed2bf8b26 100644
--- a/studio/backend/requirements/single-env/constraints.txt
+++ b/studio/backend/requirements/single-env/constraints.txt
@@ -8,13 +8,16 @@ huggingface-hub==0.36.2
 datasets==4.3.0
 pyarrow==23.0.1
 
-# FastMCP/OpenEnv compat
+# FastMCP compat
 fastmcp>=3.0.2
 mcp>=1.24,<2
 websockets>=15.0.1
 
-# anyio 4.14+ breaks cancel scope on Python 3.13 (#6483). Global cap so later
-# with-deps steps (studio.txt, data-designer-deps.txt) can't re-resolve it up.
+# Cap anyio <4.14: 4.14's new asyncio per-task cancel scope (TaskHandle/_run_coro)
+# gets exited in the wrong task on Python 3.13 under starlette's collapsing task
+# group, raising "RuntimeError: ... exit a cancel scope that isn't the current
+# task's" on streaming responses (#6483); 4.13 has no such code. Global cap so
+# later with-deps steps can't re-resolve it up.
 anyio<4.14.0
 
 pandas==2.3.3
diff --git a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
index 2cd03d8b78..56b948644f 100644
--- a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
+++ b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
@@ -3,3 +3,10 @@
 # backtrack unsloth. Relax to match the pin -- per-model 5.x routing
 # happens at runtime via the side-car venvs.
 transformers>=4.57.6
+
+# mlx-vlm / mlx-lm pull anyio>=4.14, which fights the constraints.txt cap (needed
+# for the 4.14 Python-3.13 streaming cancel-scope RuntimeError, #6483). The -c
+# constraint loses that fight on macOS-arm, leaving a half-resolved 4.14/4.13
+# anyio that also ImportErrors on TaskHandle and 500s the server. An override
+# wins the fight, so force one consistent <4.14 here too.
+anyio<4.14.0
diff --git a/studio/backend/requirements/studio.txt b/studio/backend/requirements/studio.txt
index 96fef60471..1b7f7a668c 100644
--- a/studio/backend/requirements/studio.txt
+++ b/studio/backend/requirements/studio.txt
@@ -4,13 +4,11 @@ fastapi
 uvicorn
 pydantic
 packaging
-matplotlib
+matplotlib==3.10.9
 pandas
 nest_asyncio
 datasets==4.3.0
 pyjwt
-easydict
-addict
 # gradio>=4.0.0                  # 148 MB - Studio uses React + FastAPI, not Gradio
 huggingface-hub==0.36.2
 structlog>=24.1.0
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 39240b2c27..a8a75d1cf5 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -12,6 +12,7 @@ import uuid
 from pathlib import Path
 from fastapi import APIRouter, Depends, HTTPException, Request, status
 from fastapi.responses import StreamingResponse, JSONResponse, Response
+from starlette.requests import ClientDisconnect
 from typing import Any, List, Optional, Union
 import json
 import httpx
@@ -235,8 +236,15 @@ def _sse_streaming_response(content) -> StreamingResponse:
     a one-shot connection. Two callers build their response inline instead: the
     external-provider proxy omits ``Connection: close``, and the OpenAI
     passthrough returns an empty ``keep-alive`` stream when the request is
-    cancelled before the upstream response starts."""
-    return StreamingResponse(
+    cancelled before the upstream response starts.
+
+    Built on ``_SameTaskStreamingResponse`` (not Starlette's stock
+    ``StreamingResponse``) so the SSE generator runs in the request task. The
+    legacy AnyIO task-group wrapper trips "Attempted to exit a cancel scope in a
+    different task" on Python 3.13 + httpx, which surfaced as a mid-stream
+    ``response.failed``. The streaming paths that take their response inline use
+    ``_SameTaskStreamingResponse`` directly for the same reason."""
+    return _SameTaskStreamingResponse(
         content,
         media_type = "text/event-stream",
         headers = {
@@ -750,6 +758,121 @@ def _set_stream_response_read_timeout(
         pass
 
 
+_STREAM_DISCONNECT_POLL_TIMEOUT_S = 0.25
+
+
+class _CompatSameTaskTimeout:
+    """Same-task timeout fallback for Python versions before asyncio.timeout."""
+
+    def __init__(self, timeout_s: float):
+        self.timeout_s = timeout_s
+        self._task = None
+        self._handle = None
+        self._timed_out = False
+        self._cancelling = 0
+
+    async def __aenter__(self):
+        self._task = asyncio.current_task()
+        if self._task is None:
+            return self
+        if hasattr(self._task, "cancelling"):
+            self._cancelling = self._task.cancelling()
+        loop = asyncio.get_running_loop()
+        self._handle = loop.call_later(max(self.timeout_s, 0), self._cancel_task)
+        return self
+
+    async def __aexit__(self, exc_type, exc, tb):
+        if self._handle is not None:
+            self._handle.cancel()
+        if exc_type is not None and issubclass(exc_type, asyncio.CancelledError):
+            if self._timed_out:
+                if self._task is not None and hasattr(self._task, "uncancel"):
+                    if self._task.uncancel() > self._cancelling:
+                        return None
+                raise asyncio.TimeoutError from exc
+        return None
+
+    def _cancel_task(self) -> None:
+        self._timed_out = True
+        if self._task is not None:
+            self._task.cancel()
+
+
+def _same_task_timeout(timeout_s: float):
+    timeout_ctx = getattr(asyncio, "timeout", None)
+    if timeout_ctx is not None:
+        return timeout_ctx(timeout_s)
+    return _CompatSameTaskTimeout(timeout_s)
+
+
+class _SameTaskStreamingResponse(StreamingResponse):
+    """StreamingResponse without Starlette's legacy AnyIO task-group wrapper."""
+
+    def __init__(
+        self,
+        *args,
+        unstarted_cleanup = None,
+        **kwargs,
+    ) -> None:
+        super().__init__(*args, **kwargs)
+        # Async callable invoked when the client disconnects before the body
+        # iterator is ever advanced. A generator that never started cannot run
+        # its own try/finally, so a stream that acquires resources before its
+        # first yield (the passthrough opens an upstream httpx stream eagerly)
+        # passes this to release them.
+        self._unstarted_cleanup = unstarted_cleanup
+
+    async def __call__(self, scope, receive, send) -> None:
+        # Track whether the body iterator was ever advanced: send() only emits a
+        # body message after the generator yields its first chunk, so a failure
+        # before then means it never entered its try/finally.
+        body_started = False
+
+        async def _tracking_send(message) -> None:
+            nonlocal body_started
+            if message.get("type") == "http.response.body":
+                body_started = True
+            await send(message)
+
+        try:
+            await self.stream_response(_tracking_send)
+        except OSError:
+            # Client disconnected mid-send.
+            if body_started:
+                # The generator produced at least one chunk and is suspended in
+                # its try/finally. Throw CancelledError into it (not aclose's
+                # GeneratorExit) so its `except asyncio.CancelledError` handler
+                # runs and finishes any api_monitor entry; GeneratorExit would
+                # skip it and only run `finally`. Fall back to aclose() without
+                # athrow.
+                athrow = getattr(self.body_iterator, "athrow", None)
+                if athrow is not None:
+                    try:
+                        await athrow(asyncio.CancelledError())
+                    except (asyncio.CancelledError, StopAsyncIteration, RuntimeError):
+                        pass
+                else:
+                    aclose = getattr(self.body_iterator, "aclose", None)
+                    if aclose is not None:
+                        await aclose()
+            else:
+                # http.response.start failed before the body iterator advanced,
+                # so its try/finally never armed and aclose()/athrow() are no-ops
+                # on an unstarted generator. Release any resources acquired
+                # before the first yield via the explicit cleanup hook.
+                aclose = getattr(self.body_iterator, "aclose", None)
+                if aclose is not None:
+                    await aclose()
+                if self._unstarted_cleanup is not None:
+                    try:
+                        await self._unstarted_cleanup()
+                    except Exception:
+                        pass
+            raise ClientDisconnect()
+        if self.background is not None:
+            await self.background()
+
+
 async def _aclose_stream_resources(
     *,
     watchers = (),
@@ -875,8 +998,23 @@ async def _aiter_llama_stream_items(
                     raise httpx.ReadTimeout("The model did not produce a first token in time.")
                 if response is not None:
                     _set_stream_response_read_timeout(response, remaining_s)
-                item = await asyncio.wait_for(async_iter.__anext__(), timeout = remaining_s)
+                # Keep httpx/httpcore's AnyIO cancel scope in this task.
+                # asyncio.wait_for would drive __anext__ in a child task.
+                async with _same_task_timeout(remaining_s):
+                    item = await async_iter.__anext__()
             else:
+                if (
+                    request is not None
+                    and response is not None
+                    and post_first_item_read_timeout_s is not None
+                    and last_item_at is not None
+                ):
+                    stall_remaining_s = post_first_item_read_timeout_s - (
+                        time.monotonic() - last_item_at
+                    )
+                    if stall_remaining_s <= 0:
+                        raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
+                    _set_stream_response_read_timeout(response, stall_remaining_s)
                 item = await async_iter.__anext__()
         except asyncio.TimeoutError as exc:
             if waiting_first_item:
@@ -890,6 +1028,12 @@ async def _aiter_llama_stream_items(
                 if now >= first_token_deadline:
                     raise
                 continue
+            if (
+                request is not None
+                and post_first_item_read_timeout_s is not None
+                and now - last_item_at < post_first_item_read_timeout_s
+            ):
+                continue
             raise httpx.ReadTimeout("The model stopped producing tokens mid-response.")
         if (
             last_item_at is None
@@ -1125,16 +1269,17 @@ def _detect_safetensors_features(backend, chat_template: Optional[str]) -> dict:
         model_identifier = model_id,
         log_source = "safetensors",
     )
-    # Our safetensors loop only parses <tool_call>{json}</tool_call> and
-    # <function=name>...</function>. Llama uses <|python_tag|>, Mistral uses
-    # [TOOL_CALLS]; advertising tools for those enables a pill the parser
-    # can't honour. GGUF is unaffected -- llama-server normalises every
-    # format into structured deltas.
+    # Our safetensors loop only parses <tool_call>{json}</tool_call>,
+    # <function=name>...</function>, and Gemma native <|tool_call>...<tool_call|>.
+    # Llama uses <|python_tag|>, Mistral uses [TOOL_CALLS]; advertising tools for
+    # those enables a pill the parser can't honour. GGUF is unaffected --
+    # llama-server normalises every format into structured deltas.
     if (
         flags.get("supports_tools")
         and chat_template
         and "<tool_call>" not in chat_template
         and "<function=" not in chat_template
+        and "<|tool_call>" not in chat_template
     ):
         logger.info(
             "safetensors: template advertises tools but uses an "
@@ -1297,6 +1442,24 @@ async def _await_disconnect_then_close(request, resp, cancel_event) -> None:
         return
 
 
+async def _await_disconnect_then_cancel(request, cancel_event) -> None:
+    """Set ``cancel_event`` when a same-task local stream disconnects."""
+    try:
+        while not await request.is_disconnected():
+            await asyncio.sleep(0.1)
+        cancel_event.set()
+    except asyncio.CancelledError:
+        return
+
+
+async def _stop_local_disconnect_cancel_watcher(watcher) -> None:
+    watcher.cancel()
+    try:
+        await watcher
+    except (asyncio.CancelledError, Exception):
+        pass
+
+
 # Centralized local/server tool nudge. Keep render_html guidance gated to turns
 # where the canvas tool is actually present in the tool schema; otherwise
 # small local models can hallucinate a missing tool call instead of following
@@ -1418,7 +1581,9 @@ _TOOL_XML_RE = _re.compile(
     # Hyphen in the name char-class matches MCP tool names with dashes
     # (mcp__srv__list-issues) that would otherwise leak past this strip.
     r"<(?:tool_call|function=[\w-]+)>.*?(?:</(?:tool_call|function)>|\Z)"
+    r"|<\|tool_call>.*?(?:<tool_call\|>|\Z)"
     r"|</(?:tool_call|function)>"
+    r"|<tool_call\|>"
     r"|</parameter>\s*\Z",
     _re.DOTALL,
 )
@@ -3234,7 +3399,9 @@ async def get_api_monitor_entry(entry_id: str, current_subject: str = Depends(ge
 
 @router.post("/generate/stream")
 async def generate_stream(
-    request: GenerateRequest, current_subject: str = Depends(get_current_subject)
+    request: GenerateRequest,
+    fastapi_request: Request,
+    current_subject: str = Depends(get_current_subject),
 ):
     """
     Generate a chat response with Server-Sent Events (SSE) streaming.
@@ -3284,6 +3451,13 @@ async def generate_stream(
     async def stream():
         gen = None
         completed = False
+        # Cancel the generation when the client disconnects. The generator only
+        # awaits asyncio.to_thread(next, gen, ...), so without a concurrent
+        # watcher a disconnect during a long prefill/generation would go
+        # unnoticed until the next send and the backend would keep generating.
+        disconnect_watcher = asyncio.create_task(
+            _await_disconnect_then_cancel(fastapi_request, cancel_event)
+        )
         try:
             gen = backend.generate_chat_response(
                 messages = request.messages,
@@ -3298,12 +3472,22 @@ async def generate_stream(
             )
             _DONE = object()
             while True:
+                if cancel_event.is_set():
+                    # The disconnect watcher set cancel_event between chunks.
+                    # Reset the backend here: closing the Python generator does
+                    # not signal a subprocess backend, so without this it keeps
+                    # decoding after the client is gone. The finally's reset is
+                    # guarded on cancel_event being unset, so it will not run
+                    # again for this path.
+                    backend.reset_generation_state()
+                    break
                 chunk = await asyncio.to_thread(next, gen, _DONE)
                 if chunk is _DONE:
+                    completed = True
                     break
                 yield f"data: {json.dumps({'content': chunk})}\n\n"
-            completed = True
-            yield "data: [DONE]\n\n"
+            if completed:
+                yield "data: [DONE]\n\n"
 
         except asyncio.CancelledError:
             cancel_event.set()
@@ -3315,6 +3499,7 @@ async def generate_stream(
             logger.error(f"Error during generation: {e}", exc_info = True)
             yield f"data: {json.dumps({'error': _friendly_error(e)})}\n\n"
         finally:
+            await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
             if not completed and not cancel_event.is_set():
                 cancel_event.set()
                 backend.reset_generation_state()
@@ -4742,6 +4927,9 @@ async def openai_chat_completions(
                 _tracker.__enter__()
 
                 async def audio_input_stream():
+                    disconnect_watcher = asyncio.create_task(
+                        _await_disconnect_then_cancel(request, cancel_event)
+                    )
                     try:
                         yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -4777,9 +4965,18 @@ async def openai_chat_completions(
                         api_monitor.fail(monitor_id, _friendly_error(e))
                         yield f"data: {json.dumps({'error': {'message': _friendly_error(e), 'type': 'server_error'}})}\n\n"
                     finally:
+                        await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                         _tracker.__exit__(None, None, None)
 
-                return _sse_streaming_response(audio_input_stream())
+                return _SameTaskStreamingResponse(
+                    audio_input_stream(),
+                    media_type = "text/event-stream",
+                    headers = {
+                        "Cache-Control": "no-cache",
+                        "Connection": "close",
+                        "X-Accel-Buffering": "no",
+                    },
+                )
             else:
                 try:
                     full_text = "".join(audio_input_generate())
@@ -4954,6 +5151,28 @@ async def openai_chat_completions(
         completion_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
         created = int(time.time())
 
+        def _new_chat_reasoning_extractor():
+            return _ResponsesReasoningExtractor(
+                parse_think_markers = _responses_should_parse_think_markers(
+                    payload,
+                    llama_backend,
+                )
+            )
+
+        def _gguf_chat_delta_line(delta: ChoiceDelta, finish_reason = None) -> str:
+            chunk = ChatCompletionChunk(
+                id = completion_id,
+                created = created,
+                model = model_name,
+                choices = [
+                    ChunkChoice(
+                        delta = delta,
+                        finish_reason = finish_reason,
+                    )
+                ],
+            )
+            return f"data: {chunk.model_dump_json(exclude_none = True)}\n\n"
+
         # ── Tool-calling path (agentic loop) ──────────────────
         # `_effective_enable_tools` lets `unsloth run --enable-tools/--disable-tools`
         # hard-override the per-request value, else falls back to
@@ -5066,6 +5285,9 @@ async def openai_chat_completions(
 
             async def gguf_tool_stream():
                 gen = None
+                disconnect_watcher = asyncio.create_task(
+                    _await_disconnect_then_cancel(request, cancel_event)
+                )
                 try:
                     yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -5073,9 +5295,25 @@ async def openai_chat_completions(
                     # stays free for disconnect detection.
                     gen = gguf_generate_with_tools()
                     prev_text = ""
+                    reasoning_extractor = _new_chat_reasoning_extractor()
                     _stream_usage = None
                     _stream_timings = None
                     _stream_finish = None
+
+                    def _flush_reasoning_extractor():
+                        final_reasoning, final_visible = reasoning_extractor.finish()
+                        chunks = []
+                        if final_reasoning:
+                            chunks.append(
+                                _gguf_chat_delta_line(
+                                    ChoiceDelta(reasoning_content = final_reasoning)
+                                )
+                            )
+                        if final_visible:
+                            api_monitor.append_reply(monitor_id, final_visible)
+                            chunks.append(_gguf_chat_delta_line(ChoiceDelta(content = final_visible)))
+                        return chunks
+
                     while True:
                         if cancel_event.is_set():
                             break
@@ -5094,7 +5332,10 @@ async def openai_chat_completions(
                             # cumulative cursor so the next assistant turn
                             # streams cleanly.
                             if not event["text"]:
+                                for chunk in _flush_reasoning_extractor():
+                                    yield chunk
                                 prev_text = ""
+                                reasoning_extractor = _new_chat_reasoning_extractor()
                             # Emit tool status as a custom SSE event (including
                             # empty ones to clear UI badges)
                             status_data = json.dumps(
@@ -5108,7 +5349,10 @@ async def openai_chat_completions(
 
                         if event["type"] in ("tool_start", "tool_end"):
                             if event["type"] == "tool_start":
+                                for chunk in _flush_reasoning_extractor():
+                                    yield chunk
                                 prev_text = ""
+                                reasoning_extractor = _new_chat_reasoning_extractor()
                             yield f"data: {json.dumps(event)}\n\n"
                             continue
 
@@ -5118,6 +5362,11 @@ async def openai_chat_completions(
                             _stream_finish = event.get("finish_reason")
                             continue
 
+                        if event["type"] == "reasoning_summary":
+                            # Forward server-side reasoning timing to the UI.
+                            yield f"data: {json.dumps(event)}\n\n"
+                            continue
+
                         # "content" type -- cumulative text. Sanitize the full
                         # cumulative then diff against the last sanitized
                         # snapshot so cross-chunk XML tags are handled correctly.
@@ -5130,15 +5379,33 @@ async def openai_chat_completions(
                         prev_text = clean_cumulative
                         if not new_text:
                             continue
-                        api_monitor.append_reply(monitor_id, new_text)
-                        yield _chat_content_chunk(completion_id, created, model_name, new_text)
+                        reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
+                        if reasoning_delta:
+                            yield _gguf_chat_delta_line(
+                                ChoiceDelta(reasoning_content = reasoning_delta)
+                            )
+                        if visible_delta:
+                            api_monitor.append_reply(monitor_id, visible_delta)
+                            yield _gguf_chat_delta_line(ChoiceDelta(content = visible_delta))
 
-                    yield _chat_final_chunk(
-                        completion_id,
-                        created,
-                        model_name,
-                        _clamp_finish_reason(_stream_finish),
+                    for chunk in _flush_reasoning_extractor():
+                        yield chunk
+
+                    final_chunk = ChatCompletionChunk(
+                        id = completion_id,
+                        created = created,
+                        model = model_name,
+                        choices = [
+                            ChunkChoice(
+                                delta = ChoiceDelta(),
+                                finish_reason = _clamp_finish_reason(_stream_finish),
+                            )
+                        ],
                     )
+                    # Emit the terminal chunk carrying finish_reason before the
+                    # optional usage chunk and [DONE], so OpenAI-compatible
+                    # clients can detect stop/length/tool_calls.
+                    yield f"data: {final_chunk.model_dump_json(exclude_none = True)}\n\n"
                     usage_line = _openai_stream_usage_chunk(
                         payload,
                         completion_id,
@@ -5167,6 +5434,7 @@ async def openai_chat_completions(
                     error_chunk = _openai_stream_error_chunk(e)
                     yield f"data: {json.dumps(error_chunk)}\n\n"
                 finally:
+                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                     if gen is not None:
                         try:
                             gen.close()
@@ -5174,7 +5442,15 @@ async def openai_chat_completions(
                             pass
                     _tracker.__exit__(None, None, None)
 
-            return _sse_streaming_response(gguf_tool_stream())
+            return _SameTaskStreamingResponse(
+                gguf_tool_stream(),
+                media_type = "text/event-stream",
+                headers = {
+                    "Cache-Control": "no-cache",
+                    "Connection": "close",
+                    "X-Accel-Buffering": "no",
+                },
+            )
 
         # ── Standard GGUF path (no tools) ─────────────────────
 
@@ -5210,6 +5486,9 @@ async def openai_chat_completions(
             _tracker.__enter__()
 
             async def gguf_stream_chunks():
+                disconnect_watcher = asyncio.create_task(
+                    _await_disconnect_then_cancel(request, cancel_event)
+                )
                 try:
                     yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -5217,6 +5496,7 @@ async def openai_chat_completions(
                     # stays free for disconnect detection.
                     gen = gguf_generate()
                     prev_text = ""
+                    reasoning_extractor = _new_chat_reasoning_extractor()
                     _stream_usage = None
                     _stream_timings = None
                     _stream_finish = None
@@ -5250,15 +5530,38 @@ async def openai_chat_completions(
                         prev_text = cumulative
                         if not new_text:
                             continue
-                        api_monitor.append_reply(monitor_id, new_text)
-                        yield _chat_content_chunk(completion_id, created, model_name, new_text)
+                        reasoning_delta, visible_delta = reasoning_extractor.feed(new_text)
+                        if reasoning_delta:
+                            yield _gguf_chat_delta_line(
+                                ChoiceDelta(reasoning_content = reasoning_delta)
+                            )
+                        if visible_delta:
+                            api_monitor.append_reply(monitor_id, visible_delta)
+                            yield _gguf_chat_delta_line(ChoiceDelta(content = visible_delta))
 
-                    yield _chat_final_chunk(
-                        completion_id,
-                        created,
-                        model_name,
-                        _clamp_finish_reason(_stream_finish),
+                    final_reasoning, final_visible = reasoning_extractor.finish()
+                    if final_reasoning:
+                        yield _gguf_chat_delta_line(ChoiceDelta(reasoning_content = final_reasoning))
+                    if final_visible:
+                        api_monitor.append_reply(monitor_id, final_visible)
+                        yield _gguf_chat_delta_line(ChoiceDelta(content = final_visible))
+
+                    # Final chunk
+                    final_chunk = ChatCompletionChunk(
+                        id = completion_id,
+                        created = created,
+                        model = model_name,
+                        choices = [
+                            ChunkChoice(
+                                delta = ChoiceDelta(),
+                                finish_reason = _clamp_finish_reason(_stream_finish),
+                            )
+                        ],
                     )
+                    # Emit the terminal chunk carrying finish_reason before the
+                    # optional usage chunk and [DONE], so OpenAI-compatible
+                    # clients can detect stop/length/tool_calls.
+                    yield f"data: {final_chunk.model_dump_json(exclude_none = True)}\n\n"
                     usage_line = _openai_stream_usage_chunk(
                         payload,
                         completion_id,
@@ -5285,9 +5588,18 @@ async def openai_chat_completions(
                     error_chunk = _openai_stream_error_chunk(e)
                     yield f"data: {json.dumps(error_chunk)}\n\n"
                 finally:
+                    await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                     _tracker.__exit__(None, None, None)
 
-            return _sse_streaming_response(gguf_stream_chunks())
+            return _SameTaskStreamingResponse(
+                gguf_stream_chunks(),
+                media_type = "text/event-stream",
+                headers = {
+                    "Cache-Control": "no-cache",
+                    "Connection": "close",
+                    "X-Accel-Buffering": "no",
+                },
+            )
         else:
             try:
                 # ``n`` requests several independent completions; the single
@@ -5314,14 +5626,24 @@ async def openai_chat_completions(
                             continue
                         full_text = token
 
+                    reasoning_text, visible_text = _extract_responses_reasoning(
+                        full_text,
+                        parse_think_markers = _responses_should_parse_think_markers(
+                            payload,
+                            llama_backend,
+                        ),
+                    )
+                    message_kwargs = {"content": visible_text}
+                    if reasoning_text:
+                        message_kwargs["reasoning_content"] = reasoning_text
                     _choices.append(
                         CompletionChoice(
                             index = _idx,
-                            message = CompletionMessage(content = full_text),
+                            message = CompletionMessage(**message_kwargs),
                             finish_reason = _clamp_finish_reason(completion_finish),
                         )
                     )
-                    _monitor_replies.append(full_text)
+                    _monitor_replies.append(visible_text)
                     if completion_usage:
                         # The prompt is shared across all n choices, so count its
                         # tokens ONCE (OpenAI bills only generated tokens for each
@@ -5343,7 +5665,7 @@ async def openai_chat_completions(
                         prompt_tokens_details = _prompt_tokens_details(_prompt_details),
                     ),
                 )
-                monitor_reply = full_text
+                monitor_reply = _monitor_replies[-1] if _monitor_replies else ""
                 if _n > 1:
                     monitor_reply = "\n\n".join(
                         f"Choice {_idx + 1}:\n{text}" for _idx, text in enumerate(_monitor_replies)
@@ -5553,6 +5875,9 @@ async def openai_chat_completions(
 
         async def sf_tool_stream():
             gen = None
+            disconnect_watcher = asyncio.create_task(
+                _await_disconnect_then_cancel(request, cancel_event)
+            )
             try:
                 yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -5644,6 +5969,7 @@ async def openai_chat_completions(
                 }
                 yield f"data: {json.dumps(error_chunk)}\n\n"
             finally:
+                await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                 if gen is not None:
                     try:
                         gen.close()
@@ -5652,7 +5978,15 @@ async def openai_chat_completions(
                 _sf_tracker.__exit__(None, None, None)
 
         if payload.stream:
-            return _sse_streaming_response(sf_tool_stream())
+            return _SameTaskStreamingResponse(
+                sf_tool_stream(),
+                media_type = "text/event-stream",
+                headers = {
+                    "Cache-Control": "no-cache",
+                    "Connection": "close",
+                    "X-Accel-Buffering": "no",
+                },
+            )
 
         # Non-streaming JSON: drain the loop, build one ChatCompletion.
         try:
@@ -5754,6 +6088,9 @@ async def openai_chat_completions(
         _tracker.__enter__()
 
         async def stream_chunks():
+            disconnect_watcher = asyncio.create_task(
+                _await_disconnect_then_cancel(request, cancel_event)
+            )
             try:
                 yield _chat_role_chunk(completion_id, created, model_name)
 
@@ -5830,9 +6167,18 @@ async def openai_chat_completions(
                 }
                 yield f"data: {json.dumps(error_chunk)}\n\n"
             finally:
+                await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
                 _tracker.__exit__(None, None, None)
 
-        return _sse_streaming_response(stream_chunks())
+        return _SameTaskStreamingResponse(
+            stream_chunks(),
+            media_type = "text/event-stream",
+            headers = {
+                "Cache-Control": "no-cache",
+                "Connection": "close",
+                "X-Accel-Buffering": "no",
+            },
+        )
 
     # ── Non-streaming response ────────────────────────────────────
     else:
@@ -6552,8 +6898,9 @@ def _responses_should_parse_think_markers(
     if llama_backend is not None and getattr(llama_backend, "is_loaded", False):
         if getattr(llama_backend, "reasoning_always_on", False):
             return True
-        if not getattr(llama_backend, "supports_reasoning", False):
-            return False
+        if getattr(llama_backend, "supports_reasoning", False):
+            return True
+        return False
     if chat_req.enable_thinking is True:
         return True
     return chat_req.enable_thinking is None and chat_req.reasoning_effort not in (None, "none")
@@ -6849,8 +7196,6 @@ async def _responses_non_streaming(
         # the model produced content, so clients expecting a pure tool-call turn
         # (finish_reason="tool_calls") don't see a spurious empty message item.
         output_items: list[dict] = []
-        if reasoning_text and not text and not tool_calls:
-            text = reasoning_text
         if reasoning_text:
             output_items.append(_responses_reasoning_output_item(reasoning_text))
         if text:
@@ -7163,8 +7508,8 @@ async def _responses_stream(
         client = httpx.AsyncClient(timeout = _llama_streaming_generation_timeout())
         resp = None
         lines_iter = None
-        disconnect_event = threading.Event()
         disconnect_watcher = None
+        disconnect_event = threading.Event()
         try:
             req = client.build_request(
                 "POST", target_url, json = body, headers = {"Connection": "close"}
@@ -7224,10 +7569,10 @@ async def _responses_stream(
                 )
                 return
 
+            lines_iter = resp.aiter_lines()
             disconnect_watcher = asyncio.create_task(
                 _await_disconnect_then_close(request, resp, disconnect_event)
             )
-            lines_iter = resp.aiter_lines()
             async for raw_line in _aiter_llama_stream_items(
                 lines_iter,
                 cancel_event = disconnect_event,
@@ -7347,6 +7692,7 @@ async def _responses_stream(
 
                 _apply_usage(chunk_data.get("usage"))
         except asyncio.CancelledError:
+            disconnect_event.set()
             api_monitor.finish(monitor_id, "cancelled")
             raise
         except (httpx.RemoteProtocolError, httpx.ReadError, httpx.CloseError) as e:
@@ -7413,21 +7759,6 @@ async def _responses_stream(
                     "delta": final_visible,
                 },
             )
-        if full_reasoning and not full_text and not tool_call_state:
-            for event in _ensure_message_open():
-                yield event
-            full_text = full_reasoning
-            api_monitor.set_reply(monitor_id, full_text)
-            yield _sse(
-                "response.output_text.delta",
-                {
-                    "type": "response.output_text.delta",
-                    "item_id": message_state["item_id"],
-                    "output_index": message_state["output_index"],
-                    "content_index": 0,
-                    "delta": full_text,
-                },
-            )
 
         close_items: list[tuple[int, str, dict[str, Any]]] = []
         if reasoning_state["opened"]:
@@ -7588,7 +7919,15 @@ async def _responses_stream(
         api_monitor.finish(monitor_id)
         yield _sse("response.completed", completed_response)
 
-    return _sse_streaming_response(event_generator())
+    return _SameTaskStreamingResponse(
+        event_generator(),
+        media_type = "text/event-stream",
+        headers = {
+            "Cache-Control": "no-cache",
+            "Connection": "close",
+            "X-Accel-Buffering": "no",
+        },
+    )
 
 
 @router.post("/responses")
@@ -8204,9 +8543,17 @@ async def _anthropic_tool_stream(
         drop_until_tool_end = False
 
         gen = run_gen()
+        # Concurrent disconnect watcher: the loop only polls is_disconnected()
+        # between events, so a client disconnect during a long prefill or
+        # generation step would otherwise hold the decode slot until the next
+        # event or a failed send. The watcher sets cancel_event so the backend
+        # stops promptly.
+        disconnect_watcher = asyncio.create_task(
+            _await_disconnect_then_cancel(request, cancel_event)
+        )
         try:
             while True:
-                if await request.is_disconnected():
+                if cancel_event.is_set() or await request.is_disconnected():
                     cancel_event.set()
                     return
                 event = await asyncio.to_thread(next, gen, _sentinel)
@@ -8254,6 +8601,8 @@ async def _anthropic_tool_stream(
             if _error_event is not None:
                 yield _error_event
                 return
+        finally:
+            await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
 
         stop_reason = openai_finish_to_anthropic_stop(
             captured_finish_reason, had_tool_calls = ends_on_tool_use
@@ -8290,9 +8639,17 @@ async def _anthropic_plain_stream(
         captured_finish_reason = None
 
         gen = run_gen()
+        # Concurrent disconnect watcher: the loop only polls is_disconnected()
+        # between chunks, so a client disconnect during a long prefill or
+        # generation step would otherwise hold the decode slot until the next
+        # chunk or a failed send. The watcher sets cancel_event so the backend
+        # stops promptly.
+        disconnect_watcher = asyncio.create_task(
+            _await_disconnect_then_cancel(request, cancel_event)
+        )
         try:
             while True:
-                if await request.is_disconnected():
+                if cancel_event.is_set() or await request.is_disconnected():
                     cancel_event.set()
                     return
                 cumulative = await asyncio.to_thread(next, gen, _sentinel)
@@ -8315,6 +8672,8 @@ async def _anthropic_plain_stream(
             if _error_event is not None:
                 yield _error_event
                 return
+        finally:
+            await _stop_local_disconnect_cancel_watcher(disconnect_watcher)
 
         stop_reason = openai_finish_to_anthropic_stop(captured_finish_reason, had_tool_calls = False)
         for line in emitter.finish(stop_reason = stop_reason, stop_sequence = None):
@@ -9206,7 +9565,7 @@ async def _openai_passthrough_stream(
                 except Exception:
                     pass
                 _tracker.__exit__(None, None, None)
-                return StreamingResponse(
+                return _SameTaskStreamingResponse(
                     iter(()),
                     media_type = "text/event-stream",
                     headers = {
@@ -9257,6 +9616,29 @@ async def _openai_passthrough_stream(
                 _await_disconnect_then_close(request, resp, cancel_event)
             )
             monitor_done = False
+            saw_finish_reason = False
+            saw_done = False
+            saw_stream_error = False
+            saw_tool_call_delta = False
+            last_chunk_id = completion_id
+            last_chunk_model = model_name
+            last_chunk_created = int(time.time())
+
+            def _synthetic_finish_line() -> str:
+                finish_reason = "tool_calls" if saw_tool_call_delta else "stop"
+                chunk = ChatCompletionChunk(
+                    id = last_chunk_id,
+                    created = last_chunk_created,
+                    model = last_chunk_model,
+                    choices = [
+                        ChunkChoice(
+                            delta = ChoiceDelta(),
+                            finish_reason = finish_reason,
+                        )
+                    ],
+                )
+                return f"data: {chunk.model_dump_json(exclude_none = True)}"
+
             try:
                 lines_iter = resp.aiter_lines()
                 async for raw_line in _aiter_llama_stream_items(
@@ -9270,23 +9652,117 @@ async def _openai_passthrough_stream(
                         continue
                     if not raw_line.startswith("data: "):
                         continue
+                    data_text = raw_line[6:].strip()
+                    if data_text == "[DONE]":
+                        saw_done = True
+                        if (
+                            not saw_finish_reason
+                            and not saw_stream_error
+                            and not cancel_event.is_set()
+                        ):
+                            finish_line = _synthetic_finish_line()
+                            _monitor_openai_sse_line(
+                                monitor_id,
+                                finish_line,
+                                llama_backend.context_length,
+                            )
+                            yield finish_line + "\n\n"
+                            saw_finish_reason = True
+                        _monitor_openai_sse_line(
+                            monitor_id,
+                            raw_line,
+                            llama_backend.context_length,
+                        )
+                        yield raw_line + "\n\n"
+                        monitor_done = True
+                        break
                     # Honor parallel_tool_calls=false (best-effort): drop tool_call
                     # deltas with index>=1 so only the first call streams. Only
                     # lines carrying tool_calls are reparsed; everything else is
                     # relayed byte-for-byte.
                     if payload.parallel_tool_calls is False and '"tool_calls"' in raw_line:
                         raw_line = _cap_parallel_tool_calls_sse_line(raw_line)
+                        data_text = raw_line[6:].strip()
+                    try:
+                        chunk_data = json.loads(data_text)
+                    except json.JSONDecodeError:
+                        chunk_data = None
+                    if isinstance(chunk_data, dict):
+                        if isinstance(chunk_data.get("id"), str):
+                            last_chunk_id = chunk_data["id"]
+                        if isinstance(chunk_data.get("model"), str):
+                            last_chunk_model = chunk_data["model"]
+                        if isinstance(chunk_data.get("created"), int):
+                            last_chunk_created = chunk_data["created"]
+                        choices = chunk_data.get("choices")
+                        if isinstance(choices, list) and choices:
+                            choice = choices[0]
+                            if isinstance(choice, dict):
+                                if choice.get("finish_reason"):
+                                    saw_finish_reason = True
+                                delta = choice.get("delta")
+                                if isinstance(delta, dict) and delta.get("tool_calls"):
+                                    saw_tool_call_delta = True
+                        # Detect an upstream error chunk independently of API
+                        # monitoring: when monitor_id is None (skip_api_monitor),
+                        # _monitor_openai_sse_line returns before inspecting the
+                        # error, so without this the synthetic-finish guard would
+                        # emit a successful finish_reason after a failed stream.
+                        if _monitor_openai_error_message(chunk_data):
+                            saw_stream_error = True
                     monitor_event = _monitor_openai_sse_line(
                         monitor_id,
                         raw_line,
                         llama_backend.context_length,
                     )
+                    if monitor_event == "error":
+                        saw_stream_error = True
+                    # If a trailing usage-only chunk (include_usage) arrives before
+                    # any finish chunk, emit the synthetic finish first so the order
+                    # stays finish -> usage -> [DONE], matching the other streams.
+                    if (
+                        isinstance(chunk_data, dict)
+                        and chunk_data.get("usage")
+                        and not (
+                            isinstance(chunk_data.get("choices"), list) and chunk_data["choices"]
+                        )
+                        and not saw_finish_reason
+                        and not saw_stream_error
+                        and not cancel_event.is_set()
+                    ):
+                        finish_line = _synthetic_finish_line()
+                        _monitor_openai_sse_line(
+                            monitor_id, finish_line, llama_backend.context_length
+                        )
+                        yield finish_line + "\n\n"
+                        saw_finish_reason = True
                     # Relay verbatim to preserve llama-server's native id,
                     # finish_reason, delta.tool_calls, and usage chunks.
                     yield raw_line + "\n\n"
-                    if monitor_event == "done" or raw_line[6:].strip() == "[DONE]":
+                    if monitor_event == "done":
                         monitor_done = True
                         break
+                if not saw_done and not saw_stream_error and not cancel_event.is_set():
+                    # Synthesize a finish chunk only if one was not already
+                    # emitted (e.g. before a trailing usage-only chunk), but
+                    # always close with [DONE] whenever the upstream omitted it,
+                    # so the stream ends on the [DONE] sentinel either way.
+                    if not saw_finish_reason:
+                        finish_line = _synthetic_finish_line()
+                        _monitor_openai_sse_line(
+                            monitor_id,
+                            finish_line,
+                            llama_backend.context_length,
+                        )
+                        yield finish_line + "\n\n"
+                    done_line = "data: [DONE]"
+                    _monitor_openai_sse_line(
+                        monitor_id,
+                        done_line,
+                        llama_backend.context_length,
+                    )
+                    yield done_line + "\n\n"
+                    monitor_done = True
                 if not monitor_done:
                     api_monitor.finish(
                         monitor_id,
@@ -9322,7 +9798,24 @@ async def _openai_passthrough_stream(
                 )
                 _tracker.__exit__(None, None, None)
 
-        return _sse_streaming_response(_stream())
+        async def _unstarted_cleanup() -> None:
+            # Client disconnected before the body stream started, so _stream()'s
+            # finally never ran. Release the eagerly-opened upstream resp/client
+            # and the cancel-registry entry here; the watchers and line iterator
+            # are created inside _stream(), so there is nothing else to close.
+            await _aclose_stream_resources(resp = resp, client = client)
+            _tracker.__exit__(None, None, None)
+
+        return _SameTaskStreamingResponse(
+            _stream(),
+            media_type = "text/event-stream",
+            headers = {
+                "Cache-Control": "no-cache",
+                "Connection": "close",
+                "X-Accel-Buffering": "no",
+            },
+            unstarted_cleanup = _unstarted_cleanup,
+        )
     except BaseException:
         _tracker.__exit__(None, None, None)
         raise
diff --git a/studio/backend/run.py b/studio/backend/run.py
index 709efc2098..bb57db9bcb 100644
--- a/studio/backend/run.py
+++ b/studio/backend/run.py
@@ -893,9 +893,14 @@ def _setup_server_disk_logging():
 def _cloudflare_tunnel_should_start(
     *, cloudflare: bool, host: str, secure: bool, api_only: bool, is_colab: bool
 ) -> bool:
-    """Whether to start the Cloudflare tunnel. --secure tunnels a loopback bind too;
-    non-secure keeps the 0.0.0.0-only rule. Colab/api-only never tunnel."""
-    return cloudflare and (host == "0.0.0.0" or secure) and not api_only and not is_colab
+    """Whether to start the Cloudflare tunnel. --secure exposes only the tunnel
+    (loopback bind), so it tunnels even api-only (headless secure API serving);
+    otherwise tunnel only a 0.0.0.0 bind, never api-only (Tauri) or Colab."""
+    if is_colab or not cloudflare:
+        return False
+    if secure:
+        return True
+    return host == "0.0.0.0" and not api_only
 
 
 def _apply_cli_tool_policy(enable_tools: "Optional[bool]") -> None:
@@ -919,6 +924,7 @@ def run_server(
     cloudflare: bool = True,
     secure: bool = False,
     enable_tools: "Optional[bool]" = None,
+    emit_tauri_port: bool = True,
 ):
     """
     Start the FastAPI server.
@@ -932,6 +938,9 @@ def run_server(
         llama_parallel_slots: parallel slots for llama-server
         enable_tools: explicit --enable-tools/--disable-tools policy; None leaves
             the default (tools on, per-request enable_tools honored)
+        emit_tauri_port: print the machine-readable TAURI_PORT line the desktop
+            app parses from stdout; the headless `run --api-only` path turns it
+            off so it does not pollute the documented URL/API-key banner
 
     Note:
         Signal handlers are NOT registered here so embedders (e.g. Colab) keep
@@ -974,9 +983,13 @@ def run_server(
     if _session_log is not None and not silent:
         print(f"Session log: {_session_log}")
 
-    # Set env var BEFORE importing main so CORS middleware picks it up.
+    # Set env vars BEFORE importing main so CORS middleware picks them up.
+    # secure api-only is a remote server behind Cloudflare, so it keeps the
+    # any-origin CORS profile; plain api-only stays locked to the Tauri app.
     if api_only:
         os.environ["UNSLOTH_API_ONLY"] = "1"
+    if secure:
+        os.environ["UNSLOTH_SECURE"] = "1"
 
     import nest_asyncio
 
@@ -1158,7 +1171,8 @@ def run_server(
     atexit.register(terminate_all)
 
     # Output port for Tauri (api-only), only after sockets bind and startup done.
-    if api_only:
+    # The headless `run --api-only` path opts out so it does not leak this line.
+    if api_only and emit_tauri_port:
         print(f"TAURI_PORT={port}", flush = True)
 
     # Free trycloudflare.com tunnel for 0.0.0.0 binds (the raw ip:port is often
diff --git a/studio/backend/tests/test_bypass_permissions.py b/studio/backend/tests/test_bypass_permissions.py
index 563f146816..d92509a5fe 100644
--- a/studio/backend/tests/test_bypass_permissions.py
+++ b/studio/backend/tests/test_bypass_permissions.py
@@ -135,6 +135,7 @@ def test_python_bypass_uses_bypass_preexec_and_bypass_env(captured_popen, monkey
     assert captured_popen["kwargs"]["preexec_fn"] is tools._bypass_preexec
     env = captured_popen["kwargs"]["env"]
     assert env.get("HOSTVAR") == "benign-xyz"
+    assert env.get("PYTHONIOENCODING") == "utf-8"
     assert "HF_TOKEN" not in env
 
 
@@ -151,9 +152,12 @@ def test_bash_blocklist_skipped_when_bypassed(captured_popen):
 
 
 @_POSIX_ONLY
-def test_bash_bypass_uses_bypass_preexec(captured_popen):
+def test_bash_bypass_uses_bypass_preexec(captured_popen, monkeypatch):
+    # bypass inherits benign host vars; clear so we assert _bash_exec adds none.
+    monkeypatch.delenv("PYTHONIOENCODING", raising = False)
     _bash_exec("echo hi", None, 5, "t", disable_sandbox = True)
     assert captured_popen["kwargs"]["preexec_fn"] is tools._bypass_preexec
+    assert "PYTHONIOENCODING" not in captured_popen["kwargs"]["env"]
 
 
 # ── real end-to-end python execution under bypass ───────────────────
diff --git a/studio/backend/tests/test_exec_utf8.py b/studio/backend/tests/test_exec_utf8.py
new file mode 100644
index 0000000000..90b78754ed
--- /dev/null
+++ b/studio/backend/tests/test_exec_utf8.py
@@ -0,0 +1,33 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+
+"""_python_exec must round-trip non-ASCII output end to end.
+
+Model-written code routinely contains non-ASCII (arrows, CJK, emoji). The temp
+script and the child's stdout pipe both have to be UTF-8 or it crashes/garbles
+on Windows, whose default codec is cp1252. Mirrors the report in
+unslothai/unsloth#6489. The child is ``python`` with PYTHONIOENCODING=utf-8, so
+it emits UTF-8 on every OS; this proves the round-trip on a UTF-8 host and
+guards against a regression to the OS default codec.
+"""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+_BACKEND_ROOT = Path(__file__).resolve().parents[1]
+if str(_BACKEND_ROOT) not in sys.path:
+    sys.path.insert(0, str(_BACKEND_ROOT))
+
+from core.inference.tools import _python_exec
+
+# Arrow, em-dash, accent, CJK, check mark, astral-plane emoji -- none encodable
+# in cp1252, so the OS default codec would raise on write or read.
+_UNICODE = "café — 数字 → ✓ 😀"
+
+
+@pytest.mark.parametrize("disable_sandbox", [False, True])
+def test_python_exec_round_trips_non_ascii(disable_sandbox):
+    out = _python_exec(f"print({_UNICODE!r})", disable_sandbox = disable_sandbox)
+    assert _UNICODE in out, repr(out)
diff --git a/studio/backend/tests/test_gemma_tool_parse_edge_cases.py b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
new file mode 100644
index 0000000000..8df8d37a52
--- /dev/null
+++ b/studio/backend/tests/test_gemma_tool_parse_edge_cases.py
@@ -0,0 +1,161 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Edge cases in Gemma-native tool-call parsing.
+
+Covers two failure modes:
+  1. A bare (unquoted) string argument that contains a comma, e.g.
+     ``location:New York, NY`` -- the comma must not be treated as the next
+     key boundary, or the whole call is dropped.
+  2. A tool-call marker that appears INSIDE another call's argument string is
+     data, not a real call, so it must not be promoted to a second tool call.
+"""
+
+from __future__ import annotations
+
+import json
+import sys
+from pathlib import Path
+
+_BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
+if _BACKEND_DIR not in sys.path:
+    sys.path.insert(0, _BACKEND_DIR)
+
+from core.inference.tool_call_parser import parse_tool_calls_from_text
+
+
+def _args(call: dict) -> dict:
+    return json.loads(call["function"]["arguments"])
+
+
+def test_bare_string_argument_with_comma_is_kept():
+    calls = parse_tool_calls_from_text(
+        "<|tool_call>call:get_weather{location:New York, NY,unit:celsius}<tool_call|>"
+    )
+    assert len(calls) == 1, calls
+    assert calls[0]["function"]["name"] == "get_weather"
+    assert _args(calls[0]) == {"location": "New York, NY", "unit": "celsius"}
+
+
+def test_normal_multi_key_arguments_still_split():
+    calls = parse_tool_calls_from_text('<|tool_call>call:f{a:1,b:hello,c:"x,y"}<tool_call|>')
+    assert len(calls) == 1, calls
+    # Numbers stay numeric, bare strings get quoted, an explicit quoted comma
+    # stays inside its value.
+    assert _args(calls[0]) == {"a": 1, "b": "hello", "c": "x,y"}
+
+
+def test_bare_value_with_timestamps_after_comma_is_kept():
+    # A comma followed by digits-then-colon (a timestamp/ratio) is value text,
+    # not a new key, so the whole query must be preserved as one argument.
+    calls = parse_tool_calls_from_text(
+        "<|tool_call>call:remind{query:meet at 10:00, 11:00 tomorrow,priority:high}<tool_call|>"
+    )
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"query": "meet at 10:00, 11:00 tomorrow", "priority": "high"}
+
+
+def test_marker_inside_json_argument_is_not_a_second_call():
+    # A python call whose `code` argument contains a Gemma marker string. The
+    # marker is data and must not execute as a second `terminal` call.
+    content = (
+        '<tool_call>{"name":"python","arguments":{"code":'
+        '"x = 1  # <|tool_call>call:terminal{command:ls}<tool_call|>"}}</tool_call>'
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_two_separate_gemma_calls_both_parse():
+    content = "<|tool_call>call:a{x:1}<tool_call|> and <|tool_call>call:b{y:2}<tool_call|>"
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["a", "b"], calls
+    assert _args(calls[0]) == {"x": 1}
+    assert _args(calls[1]) == {"y": 2}
+
+
+def test_mixed_format_calls_preserve_document_order():
+    # A Gemma-native call precedes a JSON-format call in the text; tools execute
+    # in returned order, so `create` must come before `read`.
+    content = (
+        "<|tool_call>call:create{path:a}<tool_call|> then "
+        '<tool_call>{"name":"read","arguments":{"path":"a"}}</tool_call>'
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["create", "read"], calls
+
+
+def test_json_marker_inside_gemma_argument_is_not_a_second_call():
+    # The reverse of the JSON-outer case: a JSON-style marker inside a Gemma
+    # call's quoted argument is code text, not a second `terminal` call.
+    content = (
+        '<|tool_call>call:python{code:<|"|>'
+        'print(<tool_call>{"name":"terminal","arguments":{"command":"ls"}}</tool_call>)'
+        '<|"|>}<tool_call|>'
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+
+
+def test_nested_gemma_marker_in_unquoted_arg_does_not_run_inner_call():
+    # An UNQUOTED Gemma value containing a literal marker: the outer object fails
+    # to normalize (the inner braces/marker break the JSON), but the inner marker
+    # is nested in the outer candidate span, so it must not be promoted to a
+    # standalone `terminal` call. The safe outcome is no executed tool call.
+    content = "<|tool_call>call:python{code:<|tool_call>call:terminal{command:ls}<tool_call|>}<tool_call|>"
+    calls = parse_tool_calls_from_text(content)
+    assert "terminal" not in [c["function"]["name"] for c in calls], calls
+
+
+def test_bare_string_array_argument_is_quoted():
+    # Gemma may emit an array of bare strings without per-element quotes; they
+    # must be quoted so the call is not dropped.
+    calls = parse_tool_calls_from_text("<|tool_call>call:label{labels:[bug,ui]}<tool_call|>")
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"labels": ["bug", "ui"]}
+
+
+def test_array_keeps_numbers_and_quoted_elements():
+    calls = parse_tool_calls_from_text(
+        '<|tool_call>call:f{nums:[1,2],tags:[<|"|>a,b<|"|>,c]}<tool_call|>'
+    )
+    assert _args(calls[0]) == {"nums": [1, 2], "tags": ["a,b", "c"]}
+
+
+def test_array_of_objects_is_normalised():
+    # Arrays of objects are a common tool-schema shape; their (unquoted) keys and
+    # bare values must be normalised too, not left verbatim, or the call drops.
+    calls = parse_tool_calls_from_text(
+        "<|tool_call>call:batch{items:[{path:a,mode:r},{path:b,mode:w}]}<tool_call|>"
+    )
+    assert len(calls) == 1, calls
+    assert _args(calls[0]) == {"items": [{"path": "a", "mode": "r"}, {"path": "b", "mode": "w"}]}
+
+
+def test_nested_array_elements_are_normalised():
+    calls = parse_tool_calls_from_text("<|tool_call>call:grid{cells:[[a,b],[c,d]]}<tool_call|>")
+    assert _args(calls[0]) == {"cells": [["a", "b"], ["c", "d"]]}
+
+
+def test_gemma_marker_inside_xml_parameter_is_not_a_second_call():
+    # An XML-style <function=...> call whose <parameter=code> value contains a
+    # Gemma marker: the marker is the parameter's data, not a separate terminal
+    # call, so only the python call must be returned.
+    content = (
+        "<tool_call><function=python><parameter=code>"
+        "x = 1  # <|tool_call>call:terminal{command:ls}<tool_call|>"
+        "</parameter></function></tool_call>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
+    assert "terminal" in _args(calls[0])["code"]
+
+
+def test_json_marker_inside_xml_parameter_is_not_a_second_call():
+    content = (
+        "<tool_call><function=python><parameter=code>"
+        'run(<tool_call>{"name":"terminal","arguments":{"command":"ls"}}</tool_call>)'
+        "</parameter></function></tool_call>"
+    )
+    calls = parse_tool_calls_from_text(content)
+    assert [c["function"]["name"] for c in calls] == ["python"], calls
diff --git a/studio/backend/tests/test_install_resolve_prebuilt.py b/studio/backend/tests/test_install_resolve_prebuilt.py
index d69eccc54d..e9941d9e62 100644
--- a/studio/backend/tests/test_install_resolve_prebuilt.py
+++ b/studio/backend/tests/test_install_resolve_prebuilt.py
@@ -209,3 +209,154 @@ def test_resolve_prebuilt_linux_amd_tooling_routes_to_fork(monkeypatch, capsys):
     out = json.loads(capsys.readouterr().out.strip().splitlines()[-1])
     assert seen["repo"] == FORK
     assert out["repo"] == FORK
+
+
+# Blackwell floor is sm_100 (data-center B100/B200, B300/GB300), below consumer
+# sm_120 -- 120 wrongly excluded data-center hosts from the prebuilt selection.
+
+
+def _gpu_linux_host(caps):
+    return _host(
+        is_linux = True,
+        is_x86_64 = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+        driver_cuda_version = (13, 1),
+        compute_caps = caps,
+    )
+
+
+def test_host_is_blackwell_includes_datacenter_parts():
+    assert ilp._host_is_blackwell(_gpu_linux_host(["10.0"])) is True  # B200 sm_100
+    assert ilp._host_is_blackwell(_gpu_linux_host(["10.3"])) is True  # B300 sm_103
+    assert ilp._host_is_blackwell(_gpu_linux_host(["12.0"])) is True  # RTX 50 sm_120
+    assert ilp._host_is_blackwell(_gpu_linux_host(["12.1"])) is True  # DGX Spark sm_121
+    assert ilp._host_is_blackwell(_gpu_linux_host(["9.0"])) is False  # Hopper
+    assert ilp._host_is_blackwell(_gpu_linux_host(["8.0"])) is False  # Ampere
+    assert ilp._host_is_blackwell(_gpu_linux_host(["9.0", "10.0"])) is True  # highest cap wins
+
+
+def _linux_cuda_artifact(runtime_line, supported_sms, min_sm, max_sm, profile):
+    return ilp.PublishedLlamaArtifact(
+        asset_name = f"app-b9739-linux-x64-{profile}.tar.gz",
+        install_kind = "linux-cuda",
+        runtime_line = runtime_line,
+        coverage_class = "newer",
+        supported_sms = supported_sms,
+        min_sm = min_sm,
+        max_sm = max_sm,
+        bundle_profile = profile,
+        rank = 50,
+    )
+
+
+def test_linux_blackwell_override_prefers_cuda13_for_datacenter(monkeypatch):
+    # Both bundles cover sm_100 and torch reports cuda12, so coverage alone can't
+    # decide -- only the sm_100 Blackwell floor lifts cuda13 to the front.
+    cuda12 = _linux_cuda_artifact(
+        "cuda12", ["86", "89", "90", "100", "120"], 86, 120, "cuda12-newer"
+    )
+    cuda13 = _linux_cuda_artifact(
+        "cuda13", ["86", "89", "90", "100", "103", "120"], 86, 120, "cuda13-newer"
+    )
+    release = ilp.PublishedReleaseBundle(
+        repo = FORK,
+        release_tag = "b9739-mix",
+        upstream_tag = "b9739",
+        assets = {cuda12.asset_name: "https://x/cuda12", cuda13.asset_name: "https://x/cuda13"},
+        artifacts = [cuda12, cuda13],
+    )
+    monkeypatch.setattr(
+        ilp,
+        "detected_linux_runtime_lines",
+        lambda: (["cuda13", "cuda12"], {"cuda13": ["/usr/lib"], "cuda12": ["/usr/lib"]}),
+    )
+
+    selection = ilp.linux_cuda_choice_from_release(
+        _gpu_linux_host(["10.0"]), release, preferred_runtime_line = "cuda12"
+    )
+    assert selection is not None
+    assert selection.primary.runtime_line == "cuda13"
+    assert selection.primary.bundle_profile == "cuda13-newer"
+
+
+def test_drop_blackwell_incapable_windows_cuda_applies_to_datacenter():
+    # B200 (sm_100) on Windows must drop the cuda-12.4 build and keep cuda13.
+    host = _host(
+        system = "Windows",
+        is_windows = True,
+        is_x86_64 = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+        compute_caps = ["10.0"],
+    )
+    cuda124 = ilp.AssetChoice(
+        repo = FORK,
+        tag = "b9739",
+        name = "llama-b9739-bin-win-cuda-12.4-x64.zip",
+        url = "https://x/124",
+        source_label = "published",
+        install_kind = "windows-cuda",
+    )
+    cuda13 = ilp.AssetChoice(
+        repo = FORK,
+        tag = "b9739",
+        name = "app-b9739-windows-x64-cuda13-newer.zip",
+        url = "https://x/13",
+        source_label = "published",
+        install_kind = "windows-cuda",
+        max_sm = 120,
+    )
+    kept = ilp._drop_blackwell_incapable_windows_cuda(host, [cuda124, cuda13])
+    assert [a.name for a in kept] == [cuda13.name]
+
+
+def test_blackwell_min_toolkit_is_sm_aware():
+    # Family floor is 12.8; sm_103/sm_121 (no native target before 12.9) lift it.
+    f = ilp._blackwell_min_toolkit_for_host
+    assert f(_gpu_linux_host(["10.0"])) == (12, 8)  # B200
+    assert f(_gpu_linux_host(["12.0"])) == (12, 8)  # RTX 50
+    assert f(_gpu_linux_host(["10.3"])) == (12, 9)  # B300
+    assert f(_gpu_linux_host(["12.1"])) == (12, 9)  # DGX Spark
+    assert f(_gpu_linux_host(["10.0", "10.3"])) == (12, 9)  # max across SMs wins
+
+
+def test_sm103_host_drops_cuda128_windows_build():
+    # B300 (sm_103) needs cuda-12.9: a legacy win-cuda-12.8 build must be dropped.
+    host = _host(
+        system = "Windows",
+        is_windows = True,
+        is_x86_64 = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+        compute_caps = ["10.3"],
+    )
+    cuda128 = ilp.AssetChoice(
+        repo = FORK,
+        tag = "b9739",
+        name = "llama-b9739-bin-win-cuda-12.8-x64.zip",
+        url = "https://x/128",
+        source_label = "published",
+        install_kind = "windows-cuda",
+    )
+    cuda129 = ilp.AssetChoice(
+        repo = FORK,
+        tag = "b9739",
+        name = "llama-b9739-bin-win-cuda-12.9-x64.zip",
+        url = "https://x/129",
+        source_label = "published",
+        install_kind = "windows-cuda",
+    )
+    kept = ilp._drop_blackwell_incapable_windows_cuda(host, [cuda128, cuda129])
+    assert [a.name for a in kept] == [cuda129.name]
+    # sm_100 stays on the 12.8 family floor and keeps the same 12.8 build.
+    b200 = _host(
+        system = "Windows",
+        is_windows = True,
+        is_x86_64 = True,
+        has_physical_nvidia = True,
+        has_usable_nvidia = True,
+        compute_caps = ["10.0"],
+    )
+    kept_b200 = ilp._drop_blackwell_incapable_windows_cuda(b200, [cuda128, cuda129])
+    assert [a.name for a in kept_b200] == [cuda128.name, cuda129.name]
diff --git a/studio/backend/tests/test_llama_cpp_tool_loop.py b/studio/backend/tests/test_llama_cpp_tool_loop.py
index 687829980b..56e028bd5a 100644
--- a/studio/backend/tests/test_llama_cpp_tool_loop.py
+++ b/studio/backend/tests/test_llama_cpp_tool_loop.py
@@ -77,6 +77,23 @@ def _tool_names(payload: dict) -> list[str]:
     ]
 
 
+def _patch_monotonic(monkeypatch, values: list[float]) -> None:
+    import core.inference.llama_cpp as llama_cpp_mod
+
+    it = iter(values)
+    last = values[-1]
+
+    def fake_monotonic() -> float:
+        nonlocal last
+        try:
+            last = next(it)
+        except StopIteration:
+            pass
+        return last
+
+    monkeypatch.setattr(llama_cpp_mod.time, "monotonic", fake_monotonic)
+
+
 def _structured_tool_call(tool_name: str, arguments: dict, call_id: str) -> list[str]:
     return [
         _sse(
@@ -200,6 +217,80 @@ def test_structured_tool_call_after_visible_preface_is_executed(monkeypatch):
     assert assistant_messages[-1]["tool_calls"][0]["function"]["name"] == "render_html"
 
 
+def test_buffered_reasoning_answer_emits_backend_summary(monkeypatch):
+    stream = [
+        _sse({"reasoning_content": "I am thinking."}),
+        _sse({"reasoning_content": " Still thinking."}),
+        _sse({"content": "Final answer."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [stream], payloads)
+    _patch_monotonic(monkeypatch, [100.0, 110.0, 172.0, 172.0])
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "answer"}],
+            tools = [{"type": "function", "function": {"name": "web_search"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    summary_index = next(
+        i for i, event in enumerate(events) if event["type"] == "reasoning_summary"
+    )
+    content_index = next(i for i, event in enumerate(events) if event["type"] == "content")
+    assert summary_index < content_index
+    assert events[summary_index]["duration_ms"] == 62000
+    assert (
+        events[content_index]["text"]
+        == "<think>I am thinking. Still thinking.</think>Final answer."
+    )
+
+
+def test_consumed_tool_final_pass_emits_latest_reasoning_summary(monkeypatch):
+    tool_stream = [
+        _sse({"reasoning_content": "Need a render."}),
+        _sse(
+            {
+                "content": '<tool_call>{"name":"render_html","arguments":{"code":"<html>ok</html>"}}</tool_call>'
+            }
+        ),
+        _done(),
+    ]
+    final_stream = [
+        _sse({"reasoning_content": "Now synthesize."}),
+        _sse({"content": "Final from tool."}),
+        _done(),
+    ]
+    payloads: list[dict] = []
+    backend = _make_backend(monkeypatch, [tool_stream, final_stream], payloads)
+    _patch_monotonic(monkeypatch, [200.0, 201.0, 203.0, 300.0, 400.0, 405.0, 405.0])
+
+    def fake_execute_tool(name, arguments, **_kwargs):
+        return "Rendered HTML canvas: Done."
+
+    monkeypatch.setattr("core.inference.tools.execute_tool", fake_execute_tool)
+
+    events = list(
+        backend.generate_chat_completion_with_tools(
+            messages = [{"role": "user", "content": "render then answer"}],
+            tools = [{"type": "function", "function": {"name": "render_html"}}],
+            max_tool_iterations = 1,
+        )
+    )
+
+    summaries = [event for event in events if event["type"] == "reasoning_summary"]
+    assert [event["duration_ms"] for event in summaries] == [2000, 5000]
+    final_summary_index = events.index(summaries[-1])
+    final_content_index = next(
+        i
+        for i, event in enumerate(events)
+        if event.get("type") == "content" and "Final from tool." in event.get("text", "")
+    )
+    assert final_summary_index < final_content_index
+
+
 def test_repeat_render_html_nudge_is_not_user_visible_error(monkeypatch):
     """A repeated render_html call is an internal no-op, not a visible card."""
 
diff --git a/studio/backend/tests/test_llama_route_timeouts.py b/studio/backend/tests/test_llama_route_timeouts.py
index 5aee6198ba..24866bd03e 100644
--- a/studio/backend/tests/test_llama_route_timeouts.py
+++ b/studio/backend/tests/test_llama_route_timeouts.py
@@ -5,6 +5,7 @@ import asyncio
 import os
 import sys
 import time
+import threading
 from types import SimpleNamespace
 
 _backend = os.path.join(os.path.dirname(__file__), "..")
@@ -40,6 +41,123 @@ def test_stream_first_item_deadline_after_headers():
     asyncio.run(_run())
 
 
+def test_stream_first_item_deadline_does_not_hop_tasks():
+    async def _run():
+        outer_task = asyncio.current_task()
+        seen_tasks = []
+
+        class _One:
+            def __init__(self):
+                self.done = False
+
+            async def __anext__(self):
+                seen_tasks.append(asyncio.current_task())
+                if self.done:
+                    raise StopAsyncIteration
+                self.done = True
+                return "data: {}"
+
+        out = []
+        async for item in inf_mod._aiter_llama_stream_items(
+            _One(),
+            first_token_deadline = time.monotonic() + 1,
+        ):
+            out.append(item)
+
+        assert out == ["data: {}"]
+        assert seen_tasks == [outer_task, outer_task]
+
+    asyncio.run(_run())
+
+
+def test_stream_first_item_deadline_uses_compat_timeout_without_task_hop(monkeypatch):
+    monkeypatch.setattr(inf_mod.asyncio, "timeout", None, raising = False)
+
+    async def _run():
+        outer_task = asyncio.current_task()
+        seen_tasks = []
+
+        class _One:
+            def __init__(self):
+                self.done = False
+
+            async def __anext__(self):
+                seen_tasks.append(asyncio.current_task())
+                if self.done:
+                    raise StopAsyncIteration
+                self.done = True
+                return "data: {}"
+
+        out = []
+        async for item in inf_mod._aiter_llama_stream_items(
+            _One(),
+            first_token_deadline = time.monotonic() + 1,
+        ):
+            out.append(item)
+
+        assert out == ["data: {}"]
+        assert seen_tasks == [outer_task, outer_task]
+
+    asyncio.run(_run())
+
+
+def test_stream_wait_stops_on_known_disconnect_before_read():
+    async def _run():
+        state = SimpleNamespace(disconnect_checks = 0)
+        cancel_event = threading.Event()
+
+        class _Request:
+            async def is_disconnected(self):
+                state.disconnect_checks += 1
+                return True
+
+        class _Unread:
+            async def __anext__(self):
+                raise AssertionError("stream should stop before reading upstream")
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _Unread(),
+            cancel_event = cancel_event,
+            request = _Request(),
+            first_token_deadline = time.monotonic() + 1,
+        ):
+            raise AssertionError("stream should stop after disconnect")
+
+        assert cancel_event.is_set()
+        assert state.disconnect_checks == 1
+
+    asyncio.run(_run())
+
+
+def test_stream_wait_does_not_shorten_upstream_read_for_disconnect_poll():
+    async def _run():
+        response = SimpleNamespace(request = SimpleNamespace(extensions = {"timeout": {}}))
+        seen_read_timeouts = []
+
+        class _Request:
+            async def is_disconnected(self):
+                return False
+
+        class _NoItem:
+            async def __anext__(self):
+                seen_read_timeouts.append(response.request.extensions["timeout"]["read"])
+                raise StopAsyncIteration
+
+        async for _ in inf_mod._aiter_llama_stream_items(
+            _NoItem(),
+            cancel_event = threading.Event(),
+            request = _Request(),
+            response = response,
+            first_token_deadline = time.monotonic() + 1,
+        ):
+            raise AssertionError("stream should end")
+
+        assert seen_read_timeouts
+        assert seen_read_timeouts[0] > inf_mod._STREAM_DISCONNECT_POLL_TIMEOUT_S
+
+    asyncio.run(_run())
+
+
 def test_preheader_send_cleanup_on_disconnect_and_cancel():
     async def _run(cancel_parent):
         state = SimpleNamespace(disconnected = False, closed = False, cancelled = False)
diff --git a/studio/backend/tests/test_mcp_servers.py b/studio/backend/tests/test_mcp_servers.py
index 90b1ade03c..12239e7113 100644
--- a/studio/backend/tests/test_mcp_servers.py
+++ b/studio/backend/tests/test_mcp_servers.py
@@ -423,6 +423,46 @@ def test_tool_healing_strip_handles_hyphenated_function_names():
     assert out == "before  after"
 
 
+def test_tool_healing_strip_handles_gemma_native_tool_call():
+    from core.tool_healing import strip_tool_call_markup
+    out = strip_tool_call_markup(
+        'before <|tool_call>call:mcp__srv__list-issues{repo:"octocat/hello"}<tool_call|> after'
+    )
+    assert out == "before  after"
+
+
+def test_tool_healing_strip_handles_gemma_close_only_marker():
+    from core.tool_healing import strip_tool_call_markup
+    assert strip_tool_call_markup("before <tool_call|> after") == "before  after"
+    assert strip_tool_call_markup("before <tool_call|> after", final = True) == "before  after"
+
+
+def test_tool_healing_parser_handles_gemma_native_windows_path():
+    from core.tool_healing import parse_tool_calls_from_text
+    import json as _json
+
+    calls = parse_tool_calls_from_text(
+        r'<|tool_call>call:ls{path:<|"|>C:\Users\wasim\repo<|"|>}<tool_call|>'
+    )
+    assert len(calls) == 1
+    assert calls[0]["function"]["name"] == "ls"
+    assert _json.loads(calls[0]["function"]["arguments"]) == {"path": r"C:\Users\wasim\repo"}
+
+
+def test_tool_healing_json_parser_preserves_literal_gemma_quote_token():
+    from core.tool_healing import parse_tool_calls_from_text
+    import json as _json
+
+    text = (
+        "<tool_call>"
+        + _json.dumps({"name": "python", "arguments": {"code": "print('<|\"|>')"}})
+        + "</tool_call>"
+    )
+    calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+    assert len(calls) == 1
+    assert _json.loads(calls[0]["function"]["arguments"]) == {"code": "print('<|\"|>')"}
+
+
 def test_gguf_allow_list_blocks_unadvertised_tool(monkeypatch):
     """A tool call not in the per-request list must be refused by the GGUF
     agentic loop (mirroring the safetensors path)."""
diff --git a/studio/backend/tests/test_message_content.py b/studio/backend/tests/test_message_content.py
new file mode 100644
index 0000000000..6da3682141
--- /dev/null
+++ b/studio/backend/tests/test_message_content.py
@@ -0,0 +1,100 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Tests for `content_to_text`, the #4383 fix for list-form message content.
+
+Loaded by file path so the test skips importing ``core.inference`` (whose
+``__init__`` pulls in the orchestrator + llama_cpp / torch).
+"""
+
+import importlib.util
+from pathlib import Path
+
+
+_BACKEND_DIR = Path(__file__).resolve().parent.parent
+
+
+def _load_message_content():
+    path = _BACKEND_DIR / "core/inference/message_content.py"
+    spec = importlib.util.spec_from_file_location("message_content_under_test", path)
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    return module
+
+
+def test_string_is_returned_unchanged():
+    mc = _load_message_content()
+    assert mc.content_to_text("hello world") == "hello world"
+    assert mc.content_to_text("") == ""
+
+
+def test_none_becomes_empty_string():
+    mc = _load_message_content()
+    assert mc.content_to_text(None) == ""
+
+
+def test_single_text_part_list():
+    mc = _load_message_content()
+    content = [{"type": "text", "text": "hello"}]
+    assert mc.content_to_text(content) == "hello"
+
+
+def test_multimodal_list_drops_non_text_parts():
+    mc = _load_message_content()
+    content = [
+        {"type": "text", "text": "describe this"},
+        {"type": "image_url", "image_url": {"url": "data:image/png;base64,AAAA"}},
+    ]
+    assert mc.content_to_text(content) == "describe this"
+
+
+def test_multiple_text_parts_joined_with_newline():
+    mc = _load_message_content()
+    content = [
+        {"type": "text", "text": "first"},
+        {"type": "text", "text": "second"},
+    ]
+    assert mc.content_to_text(content) == "first\nsecond"
+
+
+def test_bare_string_items_in_list():
+    mc = _load_message_content()
+    assert mc.content_to_text(["a", "b"]) == "a\nb"
+
+
+def test_audio_and_image_only_list_is_empty():
+    mc = _load_message_content()
+    content = [
+        {"type": "image_url", "image_url": {"url": "x"}},
+        {"type": "input_audio", "input_audio": {"data": "y", "format": "wav"}},
+    ]
+    assert mc.content_to_text(content) == ""
+
+
+def test_part_without_type_treated_as_text():
+    mc = _load_message_content()
+    # A ``text`` field with no ``type`` is treated as text.
+    assert mc.content_to_text([{"text": "untyped"}]) == "untyped"
+
+
+def test_empty_text_parts_skipped():
+    mc = _load_message_content()
+    content = [
+        {"type": "text", "text": ""},
+        {"type": "text", "text": "kept"},
+    ]
+    assert mc.content_to_text(content) == "kept"
+
+
+def test_tuple_behaves_like_list():
+    mc = _load_message_content()
+    content = ({"type": "text", "text": "x"}, {"type": "text", "text": "y"})
+    assert mc.content_to_text(content) == "x\ny"
+
+
+def test_result_supports_string_ops():
+    mc = _load_message_content()
+    # Crux of #4383: result must be a plain str for caller .strip()/.replace().
+    out = mc.content_to_text([{"type": "text", "text": "  padded  "}])
+    assert out.strip() == "padded"
+    assert isinstance(out, str)
diff --git a/studio/backend/tests/test_mlx_repair.py b/studio/backend/tests/test_mlx_repair.py
index 7e60c9c3b9..bb8dfce049 100644
--- a/studio/backend/tests/test_mlx_repair.py
+++ b/studio/backend/tests/test_mlx_repair.py
@@ -119,6 +119,68 @@ def test_repair_install_pins_transformers_and_cleans_up(monkeypatch):
     assert env.get("UV_OVERRIDE", "").endswith("overrides-darwin-arm64.txt")
 
 
+def test_install_requires_prebuilt_wheels(monkeypatch):
+    # A source distribution's PEP 517 build backend runs arbitrary code at install
+    # time, before the post-install stack check. The unattended self-heal must
+    # require pre-built wheels so a malicious resolver-selected sdist cannot execute
+    # during ordinary Studio startup. mlx/mlx-metal ship wheels only and
+    # mlx-lm/mlx-vlm publish py3-none-any wheels, so a healthy self-heal still works.
+    pytest.importorskip("transformers")
+    captured = {}
+
+    class _Result:
+        returncode = 0
+        stdout = ""
+
+    monkeypatch.setattr(mr, "_uv_executable", lambda: "/usr/bin/uv")
+    monkeypatch.setattr(
+        mr.subprocess, "run", lambda cmd, **k: captured.update(cmd = cmd) or _Result()
+    )
+    monkeypatch.setattr(mr, "mlx_stack_available", lambda: True)
+
+    assert mr.attempt_mlx_repair() is True
+    assert mr._ONLY_BINARY_ARG in captured["cmd"]
+
+
+def test_install_env_drops_secrets_and_source_redirects(monkeypatch):
+    # The unattended self-heal must not hand resolver/build code the full Studio
+    # environment: secrets and package-source redirects are dropped, while the
+    # variables uv genuinely needs are forwarded.
+    monkeypatch.setenv("HF_TOKEN", "secret-hf")
+    monkeypatch.setenv("AWS_SECRET_ACCESS_KEY", "secret-aws")
+    monkeypatch.setenv("WANDB_API_KEY", "secret-wandb")
+    monkeypatch.setenv("UV_FIND_LINKS", "/tmp/evil")
+    monkeypatch.setenv("UV_DEFAULT_INDEX", "file:///tmp/evil-index")
+    monkeypatch.setenv("UV_INDEX_URL", "https://evil.example/simple")
+    monkeypatch.setenv("PIP_INDEX_URL", "https://evil.example/simple")
+    monkeypatch.setenv("UV_CACHE_DIR", "/tmp/evil-cache")
+    monkeypatch.setenv("XDG_CACHE_HOME", "/tmp/evil-xdg-cache")
+    monkeypatch.setenv("PATH", "/usr/bin:/bin")
+    monkeypatch.setenv("HOME", "/home/studio")
+
+    env = mr._mlx_install_env()
+
+    # Secrets never reach a (potentially malicious) build/install hook.
+    for secret in ("HF_TOKEN", "AWS_SECRET_ACCESS_KEY", "WANDB_API_KEY"):
+        assert secret not in env
+    # A poisoned process env cannot repoint the install at a hostile source or
+    # an attacker-staged cache (cache poisoning / symlink writes).
+    for redirect in (
+        "UV_FIND_LINKS",
+        "UV_DEFAULT_INDEX",
+        "UV_INDEX_URL",
+        "PIP_INDEX_URL",
+        "UV_CACHE_DIR",
+        "XDG_CACHE_HOME",
+    ):
+        assert redirect not in env
+    # What uv genuinely needs is still forwarded.
+    assert env["PATH"] == "/usr/bin:/bin"
+    assert env["HOME"] == "/home/studio"
+    # UV_OVERRIDE is set by us (not inherited), so a poisoned one is ignored.
+    assert env.get("UV_OVERRIDE", "").endswith("overrides-darwin-arm64.txt")
+
+
 def test_repair_rejects_inadequate_stack(monkeypatch):
     # A successful uv run that still leaves an old/missing mlx-vlm must NOT clear
     # chat-only: attempt_mlx_repair returns False so Train/Export stay disabled.
diff --git a/studio/backend/tests/test_openai_tool_passthrough.py b/studio/backend/tests/test_openai_tool_passthrough.py
index 2586076321..aaef9e4dcc 100644
--- a/studio/backend/tests/test_openai_tool_passthrough.py
+++ b/studio/backend/tests/test_openai_tool_passthrough.py
@@ -48,6 +48,7 @@ from routes.inference import (
     _openai_passthrough_stream,
     _openai_stream_usage_chunk,
     _proxy_to_external_provider,
+    _SameTaskStreamingResponse,
     _set_or_prepend_system_message,
     openai_completions,
     openai_embeddings,
@@ -1245,6 +1246,79 @@ class TestGgufVisionToolRouting:
 
         return TestGgufVisionToolRouting._drive(_consume())
 
+    @staticmethod
+    def _sse_payloads(chunks):
+        payloads = []
+        for chunk in chunks:
+            if isinstance(chunk, bytes):
+                chunk = chunk.decode()
+            for line in str(chunk).splitlines():
+                if not line.startswith("data: "):
+                    continue
+                data = line.removeprefix("data: ")
+                if data == "[DONE]":
+                    continue
+                try:
+                    payloads.append(json.loads(data))
+                except json.JSONDecodeError:
+                    pass
+        return payloads
+
+    def _run_gguf_case(
+        self,
+        monkeypatch,
+        *,
+        generate = None,
+        tool_generate = None,
+        payload_kwargs = None,
+        backend_kwargs = None,
+    ):
+        import routes.inference as inf_mod
+
+        reset_tool_policy()
+
+        def _plain(**_kwargs):
+            raise AssertionError("plain GGUF path should not be used")
+
+        backend_data = {
+            "is_loaded": True,
+            "is_vision": False,
+            "supports_tools": tool_generate is not None,
+            "supports_reasoning": True,
+            "reasoning_always_on": True,
+            "_is_audio": False,
+            "model_identifier": "test-gguf",
+            "context_length": 4096,
+            "generate_chat_completion": generate or _plain,
+        }
+        if tool_generate is not None:
+            backend_data["generate_chat_completion_with_tools"] = tool_generate
+        if backend_kwargs:
+            backend_data.update(backend_kwargs)
+        backend = SimpleNamespace(**backend_data)
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+        monkeypatch.setattr(inf_mod, "get_llama_cpp_backend", lambda: backend)
+
+        request_data = {
+            "model": "default",
+            "messages": [{"role": "user", "content": "hi"}],
+        }
+        if payload_kwargs:
+            request_data.update(payload_kwargs)
+        payload = ChatCompletionRequest(**request_data)
+        response = self._drive(
+            openai_chat_completions(payload, request = self._Request(), current_subject = "test")
+        )
+        result = SimpleNamespace(response = response, monitor = monitor, backend = backend)
+        if request_data.get("stream"):
+            result.chunks = self._consume_response(response)
+            result.payloads = self._sse_payloads(result.chunks)
+        else:
+            result.body = json.loads(response.body)
+        return result
+
     def test_image_request_with_enabled_tools_enters_gguf_tool_loop(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1390,6 +1464,152 @@ class TestGgufVisionToolRouting:
         assert "confirm_tool_calls requires stream=true" in entry["error"]
         assert monitor.active_count() == 0
 
+    def test_standard_gguf_stream_splits_reasoning_content(self, monkeypatch):
+        def _generate(**_kwargs):
+            yield "<thi"
+            yield "<think>plan"
+            yield "<think>plan</think>vis"
+            yield "<think>plan</think>visible"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            generate = _generate,
+            payload_kwargs = {"stream": True},
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
+        assert "".join(d.get("content", "") for d in deltas) == "visible"
+        assert all("<think>" not in d.get("content", "") for d in deltas)
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible"
+
+    def test_reasoning_capable_gguf_stream_splits_reasoning_by_default(self, monkeypatch):
+        def _generate(**_kwargs):
+            yield "<think>plan</think>visible"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            generate = _generate,
+            payload_kwargs = {"stream": True},
+            backend_kwargs = {"reasoning_always_on": False},
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
+        assert "".join(d.get("content", "") for d in deltas) == "visible"
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible"
+
+    def test_reasoning_capable_gguf_stream_sanitizes_think_tags_when_disabled(self, monkeypatch):
+        def _generate(**_kwargs):
+            yield "<think>leaked</think>visible"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            generate = _generate,
+            payload_kwargs = {"stream": True, "enable_thinking": False},
+            backend_kwargs = {"reasoning_always_on": False},
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        assert "".join(d.get("reasoning_content", "") for d in deltas) == "leaked"
+        assert "".join(d.get("content", "") for d in deltas) == "visible"
+        assert all("<think>" not in d.get("content", "") for d in deltas)
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible"
+
+    def test_gguf_tool_stream_splits_reasoning_and_strips_gemma_tool_marker(self, monkeypatch):
+        def _tools(**_kwargs):
+            yield {
+                "type": "content",
+                "text": '<think>plan</think>visible <|tool_call>call:terminal{command:"ls"}<tool_call|>',
+            }
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            tool_generate = _tools,
+            payload_kwargs = {
+                "stream": True,
+                "enable_tools": True,
+                "enabled_tools": ["terminal"],
+                "messages": [{"role": "user", "content": "list files"}],
+            },
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        assert "".join(d.get("reasoning_content", "") for d in deltas) == "plan"
+        combined_content = "".join(d.get("content", "") for d in deltas)
+        assert combined_content == "visible "
+        assert "<|tool_call>" not in combined_content
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible "
+
+    def test_gguf_tool_stream_flushes_held_text_before_status_reset(self, monkeypatch):
+        def _tools(**_kwargs):
+            yield {"type": "content", "text": "answer <"}
+            yield {"type": "status", "text": ""}
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(
+            monkeypatch,
+            tool_generate = _tools,
+            payload_kwargs = {
+                "stream": True,
+                "enable_tools": True,
+                "enabled_tools": ["terminal"],
+                "messages": [{"role": "user", "content": "say literal"}],
+            },
+        )
+        deltas = [p["choices"][0].get("delta", {}) for p in result.payloads if p.get("choices")]
+
+        combined_content = "".join(d.get("content", "") for d in deltas)
+        assert combined_content == "answer <"
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "answer <"
+
+    def test_non_streaming_gguf_splits_reasoning_content(self, monkeypatch):
+        def _generate(**_kwargs):
+            yield "<think>plan</think>visible"
+            yield {
+                "type": "metadata",
+                "usage": {"prompt_tokens": 3, "completion_tokens": 2, "total_tokens": 5},
+                "finish_reason": "stop",
+            }
+
+        result = self._run_gguf_case(monkeypatch, generate = _generate)
+        body = result.body
+        message = body["choices"][0]["message"]
+
+        assert message["content"] == "visible"
+        assert message["reasoning_content"] == "plan"
+        [entry] = result.monitor.snapshot()
+        assert entry["reply"] == "visible"
+
     def test_non_streaming_gguf_n_records_all_monitor_replies(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -1552,6 +1772,61 @@ class TestApiMonitorProviderAndCompletionStreams:
         async def is_disconnected(self):
             return False
 
+    async def _run_passthrough_stream(self, monkeypatch, lines):
+        import routes.inference as inf_mod
+
+        class Request:
+            async def is_disconnected(self):
+                return False
+
+        async def fake_send(*_args, **_kwargs):
+            return httpx.Response(200, content = b"")
+
+        async def fake_items(*_args, **_kwargs):
+            for line in lines:
+                yield line
+
+        monitor = ApiMonitor(max_entries = 3)
+        monkeypatch.setattr(inf_mod, "api_monitor", monitor)
+        monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
+        monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
+        monitor_id = monitor.start(
+            endpoint = "/v1/chat/completions",
+            method = "POST",
+            model = "gguf",
+            prompt = "hi",
+        )
+        payload = ChatCompletionRequest(
+            model = "default",
+            messages = [ChatMessage(role = "user", content = "hi")],
+            stream = True,
+            tools = [
+                {
+                    "type": "function",
+                    "function": {
+                        "name": "lookup",
+                        "parameters": {"type": "object", "properties": {}},
+                    },
+                }
+            ],
+        )
+
+        response = await _openai_passthrough_stream(
+            Request(),
+            threading.Event(),
+            SimpleNamespace(
+                base_url = "http://llama.test",
+                context_length = 4096,
+                _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
+            ),
+            payload,
+            "gguf",
+            "chatcmpl-test",
+            monitor_id = monitor_id,
+        )
+        chunks = [chunk async for chunk in response.body_iterator]
+        return SimpleNamespace(chunks = chunks, body = "".join(chunks), monitor = monitor)
+
     def test_external_non_streaming_json_updates_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -1980,6 +2255,7 @@ class TestApiMonitorProviderAndCompletionStreams:
                 "chatcmpl-test",
                 monitor_id = monitor_id,
             )
+            assert isinstance(response, _SameTaskStreamingResponse)
             iterator = response.body_iterator
             first = await anext(iterator)
             assert "hello" in first
@@ -1997,6 +2273,88 @@ class TestApiMonitorProviderAndCompletionStreams:
 
         asyncio.run(_run())
 
+    def test_passthrough_stream_synthesizes_missing_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    (
+                        'data: {"id":"upstream","created":123,"model":"gguf",'
+                        '"choices":[{"index":0,"delta":{"content":"hello"}}]}'
+                    ),
+                    "data: [DONE]",
+                ],
+            )
+            body = result.body
+
+            assert '"finish_reason":"stop"' in body.replace(" ", "")
+            assert "data: [DONE]" in body
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_synthesizes_tool_call_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    (
+                        'data: {"id":"upstream","created":123,"model":"gguf",'
+                        '"choices":[{"index":0,"delta":{"tool_calls":[{"index":0,'
+                        '"id":"call_1","type":"function","function":{"name":"lookup",'
+                        '"arguments":"{}"}}]}}]}'
+                    ),
+                    "data: [DONE]",
+                ],
+            )
+            compact = result.body.replace(" ", "")
+
+            assert '"finish_reason":"tool_calls"' in compact
+            assert '"finish_reason":"stop"' not in compact
+            assert "data: [DONE]" in result.body
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_error_done_skips_synthetic_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                [
+                    'data: {"error":{"message":"boom","type":"server_error"}}',
+                    "data: [DONE]",
+                ],
+            )
+            compact = result.body.replace(" ", "")
+
+            assert '"error":{"message":"boom","type":"server_error"}' in compact
+            assert '"finish_reason"' not in compact
+            assert "data: [DONE]" in result.body
+            [entry] = result.monitor.snapshot()
+            assert entry["status"] == "error"
+            assert entry["error"] == "boom"
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
+    def test_passthrough_stream_error_eof_skips_synthetic_finish_reason(self, monkeypatch):
+        async def _run():
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                ['data: {"error":{"message":"boom","type":"server_error"}}'],
+            )
+            compact = result.body.replace(" ", "")
+
+            assert '"error":{"message":"boom","type":"server_error"}' in compact
+            assert '"finish_reason"' not in compact
+            assert "data: [DONE]" not in result.body
+            [entry] = result.monitor.snapshot()
+            assert entry["status"] == "error"
+            assert entry["error"] == "boom"
+            assert result.monitor.active_count() == 0
+
+        asyncio.run(_run())
+
     def test_passthrough_non_streaming_cancel_finalizes_monitor(self, monkeypatch):
         async def _run():
             import routes.inference as inf_mod
@@ -2058,65 +2416,20 @@ class TestApiMonitorProviderAndCompletionStreams:
 
     def test_passthrough_clean_eof_finalizes_monitor(self, monkeypatch):
         async def _run():
-            import routes.inference as inf_mod
-
-            class Request:
-                async def is_disconnected(self):
-                    return False
-
-            async def fake_send(*_args, **_kwargs):
-                return httpx.Response(200, content = b"")
-
-            async def fake_items(*_args, **_kwargs):
-                yield 'data: {"choices":[{"delta":{"content":"hello"}}]}'
-
-            monitor = ApiMonitor(max_entries = 3)
-            monkeypatch.setattr(inf_mod, "api_monitor", monitor)
-            monkeypatch.setattr(inf_mod, "_send_stream_with_preheader_cancel", fake_send)
-            monkeypatch.setattr(inf_mod, "_aiter_llama_stream_items", fake_items)
-            monitor_id = monitor.start(
-                endpoint = "/v1/chat/completions",
-                method = "POST",
-                model = "gguf",
-                prompt = "hi",
-            )
-            payload = ChatCompletionRequest(
-                model = "default",
-                messages = [ChatMessage(role = "user", content = "hi")],
-                stream = True,
-                tools = [
-                    {
-                        "type": "function",
-                        "function": {
-                            "name": "lookup",
-                            "parameters": {"type": "object", "properties": {}},
-                        },
-                    }
-                ],
+            result = await self._run_passthrough_stream(
+                monkeypatch,
+                ['data: {"choices":[{"delta":{"content":"hello"}}]}'],
             )
+            chunks = result.chunks
 
-            response = await _openai_passthrough_stream(
-                Request(),
-                threading.Event(),
-                SimpleNamespace(
-                    base_url = "http://llama.test",
-                    context_length = 4096,
-                    _request_reasoning_kwargs = lambda *_args, **_kwargs: None,
-                ),
-                payload,
-                "gguf",
-                "chatcmpl-test",
-                monitor_id = monitor_id,
-            )
-            chunks = []
-            async for chunk in response.body_iterator:
-                chunks.append(chunk)
-
-            assert chunks == ['data: {"choices":[{"delta":{"content":"hello"}}]}\n\n']
-            [entry] = monitor.snapshot()
+            assert chunks[0] == 'data: {"choices":[{"delta":{"content":"hello"}}]}\n\n'
+            compact = "".join(chunks).replace(" ", "")
+            assert '"finish_reason":"stop"' in compact
+            assert chunks[-1] == "data: [DONE]\n\n"
+            [entry] = result.monitor.snapshot()
             assert entry["status"] == "completed"
             assert entry["reply"] == "hello"
-            assert monitor.active_count() == 0
+            assert result.monitor.active_count() == 0
 
         asyncio.run(_run())
 
diff --git a/studio/backend/tests/test_responses_tool_passthrough.py b/studio/backend/tests/test_responses_tool_passthrough.py
index 0bea355668..4147746b54 100644
--- a/studio/backend/tests/test_responses_tool_passthrough.py
+++ b/studio/backend/tests/test_responses_tool_passthrough.py
@@ -59,8 +59,10 @@ from models.inference import (
     ResponsesUsage,
 )
 from routes.inference import (
+    _SameTaskStreamingResponse,
     _build_chat_request,
     _chat_tool_calls_to_responses_output,
+    _extract_responses_reasoning,
     _normalise_responses_input,
     _responses_tool_output_content,
     _responses_non_streaming,
@@ -782,6 +784,15 @@ class TestResponsesNonStreamingAdapter:
         assert "<think>" not in body["output"][1]["content"][0]["text"]
         assert "</think>" not in body["output"][1]["content"][0]["text"]
 
+    def test_unclosed_think_block_extracts_as_reasoning(self):
+        reasoning, visible = _extract_responses_reasoning(
+            "<think>partial plan",
+            parse_think_markers = True,
+        )
+
+        assert reasoning == "partial plan"
+        assert visible == ""
+
     def test_monitor_records_translated_visible_text(self, monkeypatch):
         import routes.inference as inf_mod
 
@@ -927,6 +938,38 @@ class TestResponsesNonStreamingAdapter:
         assert [item["type"] for item in body["output"]] == ["message"]
         assert body["output"][0]["content"][0]["text"] == "show <think>x</think> tags"
 
+    def test_reasoning_capable_gguf_parses_think_tags_by_default(self, monkeypatch):
+        body = self._run_with_message(
+            monkeypatch,
+            {"content": "<think>plan</think>answer"},
+            llama_backend = SimpleNamespace(
+                is_loaded = True,
+                reasoning_always_on = False,
+                supports_reasoning = True,
+            ),
+        )
+
+        assert [item["type"] for item in body["output"]] == ["reasoning", "message"]
+        assert body["output"][0]["content"] == [{"type": "reasoning_text", "text": "plan"}]
+        assert body["output"][1]["content"][0]["text"] == "answer"
+
+    def test_reasoning_capable_gguf_sanitizes_think_tags_when_disabled(self, monkeypatch):
+        payload = ResponsesRequest(input = "hi", reasoning = {"effort": "none"})
+        body = self._run_with_message(
+            monkeypatch,
+            {"content": "<think>leaked</think>answer"},
+            payload = payload,
+            llama_backend = SimpleNamespace(
+                is_loaded = True,
+                reasoning_always_on = False,
+                supports_reasoning = True,
+            ),
+        )
+
+        assert [item["type"] for item in body["output"]] == ["reasoning", "message"]
+        assert body["output"][0]["content"] == [{"type": "reasoning_text", "text": "leaked"}]
+        assert body["output"][1]["content"][0]["text"] == "answer"
+
     def test_structured_reasoning_content_extracts_text_parts(self, monkeypatch):
         body = self._run_with_message(
             monkeypatch,
@@ -949,7 +992,7 @@ class TestResponsesNonStreamingAdapter:
         assert [item["type"] for item in body["output"]] == ["message"]
         assert body["output"][0]["content"][0]["text"] == "33"
 
-    def test_reasoning_only_is_also_visible_message_text(self, monkeypatch):
+    def test_reasoning_only_stays_out_of_visible_message_text(self, monkeypatch):
         payload = ResponsesRequest(input = "hi", reasoning = {"effort": "high"})
         body = self._run_with_message(
             monkeypatch,
@@ -957,9 +1000,8 @@ class TestResponsesNonStreamingAdapter:
             payload = payload,
         )
 
-        assert [item["type"] for item in body["output"]] == ["reasoning", "message"]
+        assert [item["type"] for item in body["output"]] == ["reasoning"]
         assert body["output"][0]["content"][0]["text"] == "plan"
-        assert body["output"][1]["content"][0]["text"] == "plan"
 
 
 # =====================================================================
@@ -1033,6 +1075,36 @@ class TestResponsesStreamAdapter:
             ),
         )
 
+    def test_stream_response_avoids_legacy_receive_watcher(self, monkeypatch):
+        self._install_stream_mock(
+            monkeypatch,
+            [{"choices": [{"delta": {"content": "33"}}]}],
+        )
+        payload = ResponsesRequest(input = "hi", stream = True)
+        messages = [ChatMessage(role = "user", content = "hi")]
+
+        async def run():
+            response = await _responses_stream(payload, messages, self._Request())
+            assert isinstance(response, _SameTaskStreamingResponse)
+
+            sent = []
+
+            async def receive():
+                raise AssertionError("Responses streams poll disconnects in the generator")
+
+            async def send(message):
+                sent.append(message)
+
+            await response({"type": "http", "asgi": {"spec_version": "2.3"}}, receive, send)
+            return sent
+
+        sent = asyncio.run(run())
+
+        assert sent[0]["type"] == "http.response.start"
+        body = b"".join(message.get("body", b"") for message in sent).decode()
+        assert "response.output_text.delta" in body
+        assert '"delta":"33"' in body.replace(" ", "")
+
     def test_split_think_markers_stream_as_reasoning_and_visible_text(self, monkeypatch):
         chunks = [
             {"choices": [{"delta": {"content": "<thi"}}]},
@@ -1286,7 +1358,7 @@ class TestResponsesStreamAdapter:
         assert entry["status"] == "completed"
         assert entry["reply"] == "tail"
 
-    def test_reasoning_only_fallback_updates_monitor(self, monkeypatch):
+    def test_reasoning_only_stream_does_not_update_visible_monitor_reply(self, monkeypatch):
         import routes.inference as inf_mod
 
         class FakeExtractor:
@@ -1327,15 +1399,16 @@ class TestResponsesStreamAdapter:
 
         lines = asyncio.run(run())
 
-        assert self._payloads(lines, "response.output_text.delta")[-1]["delta"] == "plan"
+        assert self._payloads(lines, "response.output_text.delta") == []
+        assert self._payloads(lines, "response.reasoning_text.delta")[-1]["delta"] == "plan"
         [entry] = monitor.snapshot()
         assert entry["status"] == "completed"
-        assert entry["reply"] == "plan"
+        assert entry["reply"] == ""
 
-    def test_literal_think_tags_stream_as_visible_text_without_reasoning_request(self, monkeypatch):
+    def test_reasoning_capable_gguf_stream_parses_think_tags_by_default(self, monkeypatch):
         chunks = [
-            {"choices": [{"delta": {"content": "show <thi"}}]},
-            {"choices": [{"delta": {"content": "nk>x</think> tags"}}]},
+            {"choices": [{"delta": {"content": "<thi"}}]},
+            {"choices": [{"delta": {"content": "nk>plan</think>answer"}}]},
             {"choices": [], "usage": {"prompt_tokens": 2, "completion_tokens": 3}},
         ]
         self._install_stream_mock(monkeypatch, chunks)
@@ -1350,13 +1423,15 @@ class TestResponsesStreamAdapter:
 
         reasoning_deltas = self._payloads(lines, "response.reasoning_text.delta")
         text_deltas = self._payloads(lines, "response.output_text.delta")
-        assert reasoning_deltas == []
-        assert "".join(event["delta"] for event in text_deltas) == "show <think>x</think> tags"
+        assert "".join(event["delta"] for event in reasoning_deltas) == "plan"
+        assert "".join(event["delta"] for event in text_deltas) == "answer"
         completed = self._payloads(lines, "response.completed")[0]
-        assert [item["type"] for item in completed["response"]["output"]] == ["message"]
-        assert completed["response"]["output"][0]["content"][0]["text"] == (
-            "show <think>x</think> tags"
-        )
+        assert [item["type"] for item in completed["response"]["output"]] == [
+            "reasoning",
+            "message",
+        ]
+        assert completed["response"]["output"][0]["content"][0]["text"] == "plan"
+        assert completed["response"]["output"][1]["content"][0]["text"] == "answer"
 
     def test_non_reasoning_gguf_stream_keeps_literal_think_tags_visible(self, monkeypatch):
         chunks = [
@@ -1384,7 +1459,7 @@ class TestResponsesStreamAdapter:
             "show <think>x</think> tags"
         )
 
-    def test_reasoning_only_streams_as_visible_message_text(self, monkeypatch):
+    def test_reasoning_only_stream_stays_out_of_visible_message_text(self, monkeypatch):
         chunks = [
             {"choices": [{"delta": {"content": "<think>plan</think>"}}]},
             {"choices": [], "usage": {"prompt_tokens": 2, "completion_tokens": 3}},
@@ -1402,14 +1477,34 @@ class TestResponsesStreamAdapter:
         reasoning_deltas = self._payloads(lines, "response.reasoning_text.delta")
         text_deltas = self._payloads(lines, "response.output_text.delta")
         assert "".join(event["delta"] for event in reasoning_deltas) == "plan"
-        assert "".join(event["delta"] for event in text_deltas) == "plan"
+        assert text_deltas == []
         completed = self._payloads(lines, "response.completed")[0]
-        assert [item["type"] for item in completed["response"]["output"]] == [
-            "reasoning",
-            "message",
-        ]
+        assert [item["type"] for item in completed["response"]["output"]] == ["reasoning"]
+        assert completed["response"]["output"][0]["content"][0]["text"] == "plan"
+
+    def test_unclosed_think_stream_stays_out_of_visible_message_text(self, monkeypatch):
+        chunks = [
+            {"choices": [{"delta": {"content": "<thi"}}]},
+            {"choices": [{"delta": {"content": "nk>plan"}}]},
+            {"choices": [], "usage": {"prompt_tokens": 2, "completion_tokens": 3}},
+        ]
+        self._install_stream_mock(monkeypatch, chunks)
+        payload = ResponsesRequest(input = "hi", stream = True, reasoning = {"effort": "high"})
+        messages = [ChatMessage(role = "user", content = "hi")]
+
+        async def run():
+            response = await _responses_stream(payload, messages, self._Request())
+            return await self._collect(response)
+
+        lines = asyncio.run(run())
+
+        reasoning_deltas = self._payloads(lines, "response.reasoning_text.delta")
+        text_deltas = self._payloads(lines, "response.output_text.delta")
+        assert "".join(event["delta"] for event in reasoning_deltas) == "plan"
+        assert text_deltas == []
+        completed = self._payloads(lines, "response.completed")[0]
+        assert [item["type"] for item in completed["response"]["output"]] == ["reasoning"]
         assert completed["response"]["output"][0]["content"][0]["text"] == "plan"
-        assert completed["response"]["output"][1]["content"][0]["text"] == "plan"
 
     def test_structured_reasoning_content_streams_as_reasoning(self, monkeypatch):
         chunks = [
diff --git a/studio/backend/tests/test_safetensors_capability_advertise.py b/studio/backend/tests/test_safetensors_capability_advertise.py
index 13cb6bbd46..671af93708 100644
--- a/studio/backend/tests/test_safetensors_capability_advertise.py
+++ b/studio/backend/tests/test_safetensors_capability_advertise.py
@@ -203,6 +203,20 @@ def test_detect_safetensors_features_function_xml_format_keeps_tools_on():
     assert flags["supports_tools"] is True
 
 
+def test_detect_safetensors_features_gemma_native_tool_call_keeps_tools_on():
+    """Gemma 4 emits <|tool_call>call:name{...}<tool_call|>, which the shared
+    parser now reads, so the gate must not suppress tools for it."""
+    from routes.inference import _detect_safetensors_features
+
+    tpl_with_gemma_native = (
+        "{%- if tools -%}Tool call format: "
+        "<|tool_call>call:name{key:value}<tool_call|>{%- endif -%}"
+    )
+    backend = SimpleNamespace(active_model_name = "unsloth/gemma-4-12b-it")
+    flags = _detect_safetensors_features(backend, tpl_with_gemma_native)
+    assert flags["supports_tools"] is True
+
+
 # Qwen3.5 family pin: the live GGUF + safetensors templates both wrap tool
 # calls as ``<tool_call>\n<function=name>...``. Faithful slice so the
 # classifier never silently regresses for this family.
diff --git a/studio/backend/tests/test_safetensors_tool_loop.py b/studio/backend/tests/test_safetensors_tool_loop.py
index d61c0b389c..3f2d49f0dd 100644
--- a/studio/backend/tests/test_safetensors_tool_loop.py
+++ b/studio/backend/tests/test_safetensors_tool_loop.py
@@ -10,6 +10,7 @@ calls, tool-result feedback, bad-JSON heal, duplicate-call short-circuit,
 ``__IMAGES__`` sentinel stripping, executor errors, cancel, and the iteration cap.
 """
 
+import json
 import threading
 from typing import cast
 
@@ -62,6 +63,51 @@ class TestParser:
         assert parse_tool_calls_from_text(text)[0]["function"]["name"] == "python"
         assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
 
+    def test_gemma_native_tool_call(self):
+        text = '<|tool_call>call:terminal{command:"ls -la",workdir:"."}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "terminal"
+        args = json.loads(result[0]["function"]["arguments"])
+        assert args == {"command": "ls -la", "workdir": "."}
+
+    def test_gemma_native_tool_call_template_quotes(self):
+        text = '<|tool_call>call:web_search{query:<|"|>openai news<|"|>}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "web_search"
+        assert json.loads(result[0]["function"]["arguments"]) == {"query": "openai news"}
+
+    def test_gemma_native_tool_call_template_quotes_escape_backslashes(self):
+        text = r'<|tool_call>call:ls{path:<|"|>C:\Users\wasim\repo<|"|>}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "ls"
+        assert json.loads(result[0]["function"]["arguments"]) == {"path": r"C:\Users\wasim\repo"}
+
+    def test_gemma_native_tool_call_hyphenated_argument_name(self):
+        text = '<|tool_call>call:mcp__srv__create-issue{issue-title:"Bug report"}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "mcp__srv__create-issue"
+        assert json.loads(result[0]["function"]["arguments"]) == {"issue-title": "Bug report"}
+
+    def test_gemma_native_tool_call_keeps_braces_inside_string_value(self):
+        text = '<|tool_call>call:terminal{command:"echo {foo:bar}"}<tool_call|>'
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert result[0]["function"]["name"] == "terminal"
+        assert json.loads(result[0]["function"]["arguments"]) == {"command": "echo {foo:bar}"}
+
+    def test_gemma_native_tool_call_bare_string_values(self):
+        text = "<|tool_call>call:get_weather{location:Tokyo,unit:celsius}<tool_call|>"
+        result = parse_tool_calls_from_text(text)
+        assert len(result) == 1
+        assert json.loads(result[0]["function"]["arguments"]) == {
+            "location": "Tokyo",
+            "unit": "celsius",
+        }
+
     def test_xml_function_call(self):
         text = "<function=python><parameter=code>print('hi')</parameter></function>"
         result = parse_tool_calls_from_text(text)
@@ -121,6 +167,7 @@ class TestParser:
 
     def test_has_tool_signal(self):
         assert has_tool_signal("blah <tool_call> x")
+        assert has_tool_signal("blah <|tool_call>call:terminal")
         assert has_tool_signal("hi <function=foo>...")
         assert not has_tool_signal("hello world")
 
@@ -139,6 +186,8 @@ class TestParser:
     def test_strip_markup_closed(self):
         text = "before <tool_call>{}</tool_call> after"
         assert strip_tool_markup(text) == "before  after"
+        text = 'before <|tool_call>call:terminal{command:"ls"}<tool_call|> after'
+        assert strip_tool_markup(text) == "before  after"
 
     def test_strip_markup_unclosed_final(self):
         text = "before <tool_call>{partial"
@@ -146,6 +195,7 @@ class TestParser:
         assert strip_tool_markup(text, final = True) == "before"
         # Without final=True the unclosed run is preserved.
         assert "partial" in strip_tool_markup(text)
+        assert strip_tool_markup("before <|tool_call>call:terminal{", final = True) == "before"
 
     def test_streaming_strip_respects_disabled_healing(self):
         raw = 'before <tool_call>{"name":"web_search"'
diff --git a/studio/backend/tests/test_secure_tunnel_gate.py b/studio/backend/tests/test_secure_tunnel_gate.py
index d254121c14..bf9836e8f3 100644
--- a/studio/backend/tests/test_secure_tunnel_gate.py
+++ b/studio/backend/tests/test_secure_tunnel_gate.py
@@ -31,11 +31,14 @@ from run import _cloudflare_tunnel_should_start as should_start  # noqa: E402
         # --no-cloudflare always wins.
         (False, "0.0.0.0", False, False, False, False),
         (False, "127.0.0.1", True, False, False, False),
-        # api-only and Colab never tunnel.
+        # Non-secure api-only never tunnels (Tauri).
         (True, "0.0.0.0", False, True, False, False),
-        (True, "127.0.0.1", True, True, False, False),
+        # --secure tunnels even api-only (headless secure API server).
+        (True, "127.0.0.1", True, True, False, True),
+        # Colab never tunnels, even --secure.
         (True, "0.0.0.0", False, False, True, False),
         (True, "127.0.0.1", True, False, True, False),
+        (True, "127.0.0.1", True, True, True, False),
     ],
 )
 def test_cloudflare_gate(cloudflare, host, secure, api_only, is_colab, expected):
@@ -162,3 +165,46 @@ def test_failclosed_message_present_in_source():
         "A secure Cloudflare link is not allowed, use --no-secure which provides a 0.0.0.0 link"
         in src
     )
+
+
+@pytest.mark.parametrize(
+    "api_only,secure,expected",
+    [
+        (False, False, ["*"]),  # plain server: any origin
+        (False, True, ["*"]),  # secure UI server: any origin
+        (True, True, ["*"]),  # secure api-only: remote browsers need any origin
+        (True, False, "tauri"),  # local api-only: locked to the Tauri app
+    ],
+)
+def test_cors_origins_for_mode(api_only, secure, expected):
+    from utils.host_policy import cors_origins_for_mode
+    origins = cors_origins_for_mode(api_only = api_only, secure = secure)
+    if expected == "tauri":
+        assert origins != ["*"] and any(o.startswith("tauri://") for o in origins)
+    else:
+        assert origins == expected
+
+
+def test_run_server_exports_secure_env_for_cors():
+    # run_server must export UNSLOTH_SECURE before importing main so the CORS
+    # profile can tell remote secure serving from local Tauri use.
+    src = (_BACKEND / "run.py").read_text(encoding = "utf-8")
+    assert 'os.environ["UNSLOTH_SECURE"] = "1"' in src
+
+
+def test_run_server_emit_tauri_port_defaults_on():
+    # Default on keeps the desktop app's stdout contract; the headless
+    # `run --api-only` path opts out explicitly.
+    import inspect
+
+    import run
+
+    params = inspect.signature(run.run_server).parameters
+    assert "emit_tauri_port" in params
+    assert params["emit_tauri_port"].default is True
+
+
+def test_tauri_port_print_is_gated_in_source():
+    # The TAURI_PORT line must depend on emit_tauri_port, not api_only alone.
+    src = (_BACKEND / "run.py").read_text(encoding = "utf-8")
+    assert "if api_only and emit_tauri_port:" in src
diff --git a/studio/backend/tests/test_ssm_runtime.py b/studio/backend/tests/test_ssm_runtime.py
index 2f6bae9b79..bb0caa2887 100644
--- a/studio/backend/tests/test_ssm_runtime.py
+++ b/studio/backend/tests/test_ssm_runtime.py
@@ -445,20 +445,40 @@ def test_pre_import_gate_is_transformers_free():
     import utils.security.file_security as fs
     import utils.security.consent as consent
 
-    for m in list(_sys.modules):
-        if m == "transformers" or m.startswith("transformers.") or m == "utils.models.model_config":
+    def _is_gated_module(name: str) -> bool:
+        return (
+            name == "transformers"
+            or name.startswith("transformers.")
+            or name == "utils.models.model_config"
+        )
+
+    # Snapshot then remove the modules so we can assert the gate does not re-import them.
+    # Restore the originals afterwards (finally): popping utils.models.model_config without
+    # restoring it makes a later importer get a fresh instance, so tests that patched the
+    # first instance (e.g. test_vision_cache) miss and hit the real network path.
+    _saved = {m: _sys.modules[m] for m in list(_sys.modules) if _is_gated_module(m)}
+    for m in _saved:
+        _sys.modules.pop(m, None)
+
+    try:
+        with patch.object(fs, "_fetch_security_status", return_value = None):
+            fs.evaluate_file_security("nvidia/Nemotron-H-8B", load_subdirs = ())
+        with patch.object(
+            consent, "_load_remote_code_configs", return_value = [{"model_type": "nemotron_h"}]
+        ):
+            from utils.security import evaluate_remote_code_consent_for_targets
+            evaluate_remote_code_consent_for_targets(
+                ["nvidia/Nemotron-H-8B"], trust_remote_code = True
+            )
+
+        assert "transformers" not in _sys.modules
+        assert "utils.models.model_config" not in _sys.modules
+    finally:
+        # Drop anything the gate imported, then rebind the original module objects so later
+        # tests see the same instances they captured at import time.
+        for m in [m for m in list(_sys.modules) if _is_gated_module(m) and m not in _saved]:
             _sys.modules.pop(m, None)
-
-    with patch.object(fs, "_fetch_security_status", return_value = None):
-        fs.evaluate_file_security("nvidia/Nemotron-H-8B", load_subdirs = ())
-    with patch.object(
-        consent, "_load_remote_code_configs", return_value = [{"model_type": "nemotron_h"}]
-    ):
-        from utils.security import evaluate_remote_code_consent_for_targets
-        evaluate_remote_code_consent_for_targets(["nvidia/Nemotron-H-8B"], trust_remote_code = True)
-
-    assert "transformers" not in _sys.modules
-    assert "utils.models.model_config" not in _sys.modules
+        _sys.modules.update(_saved)
 
 
 def test_pre_import_gate_skips_subdir_computation():
diff --git a/studio/backend/tests/test_tool_call_parser_strict.py b/studio/backend/tests/test_tool_call_parser_strict.py
index 8ff41342d7..931d8a705d 100644
--- a/studio/backend/tests/test_tool_call_parser_strict.py
+++ b/studio/backend/tests/test_tool_call_parser_strict.py
@@ -106,6 +106,54 @@ class TestParityWithJsonStyle:
         assert json.loads(js[0]["function"]["arguments"]) == {"query": q}
 
 
+class TestGemmaNativeStyle:
+    def test_closed_native_call_with_trailing_prose_is_accepted(self):
+        text = (
+            '<|tool_call>call:terminal{command:"ls -la",workdir:"."}<tool_call|>' " running it now"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "terminal"
+        assert json.loads(calls[0]["function"]["arguments"]) == {
+            "command": "ls -la",
+            "workdir": ".",
+        }
+
+    def test_unclosed_native_call_requires_healing(self):
+        text = '<|tool_call>call:terminal{command:"ls"}'
+        assert parse_tool_calls_from_text(text, allow_incomplete = False) == []
+        calls = parse_tool_calls_from_text(text, allow_incomplete = True)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "terminal"
+
+    def test_hyphenated_native_argument_name_is_accepted(self):
+        text = '<|tool_call>call:mcp__srv__create-issue{issue-title:"Bug report"}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert calls[0]["function"]["name"] == "mcp__srv__create-issue"
+        assert json.loads(calls[0]["function"]["arguments"]) == {"issue-title": "Bug report"}
+
+    def test_native_template_quotes_preserve_windows_path(self):
+        text = r'<|tool_call>call:ls{path:<|"|>C:\Users\wasim\repo<|"|>}<tool_call|>'
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert json.loads(calls[0]["function"]["arguments"]) == {"path": r"C:\Users\wasim\repo"}
+
+    def test_bare_unquoted_string_values_are_accepted(self):
+        # Gemma can emit enum/string args unquoted; bare JSON scalars stay typed.
+        text = (
+            "<|tool_call>call:get_weather{location:Tokyo,unit:celsius,days:3,live:true}<tool_call|>"
+        )
+        calls = parse_tool_calls_from_text(text, allow_incomplete = False)
+        assert len(calls) == 1
+        assert json.loads(calls[0]["function"]["arguments"]) == {
+            "location": "Tokyo",
+            "unit": "celsius",
+            "days": 3,
+            "live": True,
+        }
+
+
 class TestHealingPathUnaffected:
     def test_auto_heal_still_repairs_unclosed_function(self):
         text = "<function=web_search><parameter=query>cats"
diff --git a/studio/backend/tests/test_tool_xml_strip.py b/studio/backend/tests/test_tool_xml_strip.py
index 2ba3310fbe..c2dc1fe8db 100644
--- a/studio/backend/tests/test_tool_xml_strip.py
+++ b/studio/backend/tests/test_tool_xml_strip.py
@@ -125,6 +125,14 @@ def test_strips_orphan_closing_tag():
     # Mid-string </parameter> intentionally preserved (see preserve test).
 
 
+def test_strips_gemma_native_orphan_closing_tag():
+    cleaned = _TOOL_XML_RE.sub("", "Tool call drained.<tool_call|>Visible tail.")
+
+    assert "<tool_call|>" not in cleaned
+    assert "Tool call drained." in cleaned
+    assert "Visible tail." in cleaned
+
+
 # ── Tail-only </parameter> (PR #5735 follow-up) ───────────────────
 
 
diff --git a/studio/backend/tests/test_torchao_select.py b/studio/backend/tests/test_torchao_select.py
index 393a74daaf..a99eb4c45c 100644
--- a/studio/backend/tests/test_torchao_select.py
+++ b/studio/backend/tests/test_torchao_select.py
@@ -69,3 +69,14 @@ def test_default_spec_matches_table(monkeypatch):
     mod = _load_module(monkeypatch)
     assert mod._TORCHAO_DEFAULT_SPEC == "torchao==0.14.0"
     assert mod._select_torchao_spec("2.9.0") == mod._TORCHAO_DEFAULT_SPEC
+
+
+def test_skips_torchao_on_windows_rocm():
+    """The overrides step must skip torchao on Windows ROCm: no working build exists
+    there (it imports an absent c10d backend and crashes transformers.quantizers),
+    so the installer skips it and relies on the runtime stub instead."""
+    source = _INSTALL_SCRIPT.read_text(encoding = "utf-8")
+    # Branches on the Windows-ROCm marker set by _ensure_rocm_torch ...
+    assert "elif _rocm_windows_torch_installed:" in source
+    # ... and reports the skip in the progress label.
+    assert "dependency overrides (skipped, Windows ROCm)" in source
diff --git a/studio/backend/tests/test_vision_cache.py b/studio/backend/tests/test_vision_cache.py
index 192bec53c8..4349687c54 100644
--- a/studio/backend/tests/test_vision_cache.py
+++ b/studio/backend/tests/test_vision_cache.py
@@ -41,8 +41,20 @@ from utils.models.model_config import (
 
 
 @pytest.fixture(autouse = True)
-def _clear_vision_cache():
-    """Ensure every test starts with a fresh cache."""
+def _clear_vision_cache(tmp_path, monkeypatch):
+    """Ensure every test starts with a fresh cache, from an empty working dir.
+
+    ``is_vision_model`` calls ``is_local_path`` first: any relative model id that
+    happens to exist on disk (``Path(name).exists()``) is treated as a local
+    model, short-circuiting before the mocked detection internals run. The CI cwd
+    (``studio/backend``) and the HF cache can contain dirs whose names collide
+    with the synthetic remote ids used here (``org/my-vlm``, ``model-a``,
+    ``broken/model`` ...), which made these tests fail with "called 0 times".
+    Running each test from a fresh empty ``tmp_path`` removes that collision
+    while leaving the real ``is_local_path`` logic intact (the local-GGUF tests
+    pass absolute ``tmp_path`` paths, unaffected by cwd).
+    """
+    monkeypatch.chdir(tmp_path)
     _vision_detection_cache.clear()
     yield
     _vision_detection_cache.clear()
diff --git a/studio/backend/utils/host_policy.py b/studio/backend/utils/host_policy.py
index bd9ebd68ba..f506eadc03 100644
--- a/studio/backend/utils/host_policy.py
+++ b/studio/backend/utils/host_policy.py
@@ -34,6 +34,26 @@ def is_external_host(host: str) -> bool:
     return host.lower() not in _LOOPBACK_HOSTS
 
 
+# Tauri desktop webview origins. api-only serving (the desktop app calling a
+# local backend) locks CORS to these.
+_TAURI_CORS_ORIGINS = (
+    "tauri://localhost",  # Linux/macOS Tauri webview
+    "http://tauri.localhost",  # Windows Tauri webview
+    "http://localhost",  # dev fallback
+    "http://localhost:5173",  # Tauri dev/Vite
+    "http://127.0.0.1:5173",  # Tauri dev/Vite fallback
+)
+
+
+def cors_origins_for_mode(*, api_only: bool, secure: bool) -> list[str]:
+    """Allowed CORS origins. Default is any-origin (["*"]); api-only locks down
+    to the Tauri desktop app, except in secure mode where the API is published
+    over Cloudflare and must stay reachable from remote browser origins."""
+    if api_only and not secure:
+        return list(_TAURI_CORS_ORIGINS)
+    return ["*"]
+
+
 def apply_stdio_mcp_loopback_default(host: str, *, is_colab: bool = False) -> None:
     """Default stdio MCP servers on when bound to loopback.
 
diff --git a/studio/backend/utils/mlx_repair.py b/studio/backend/utils/mlx_repair.py
index 06d0289166..7e980deb82 100644
--- a/studio/backend/utils/mlx_repair.py
+++ b/studio/backend/utils/mlx_repair.py
@@ -49,6 +49,56 @@ MLX_PACKAGES = tuple(f"{name}>={version}" for name, version in _MLX_MIN_VERSIONS
 _MLX_REINSTALL_ARGS = tuple(
     arg for name in _MLX_PACKAGE_NAMES for arg in ("--reinstall-package", name)
 )
+# Require pre-built wheels for the unattended self-heal. A source distribution's
+# PEP 517 build backend runs arbitrary code at install time, and this install is
+# default-on, resolver-driven, and runs before the post-install stack check can
+# reject anything. mlx/mlx-metal ship wheels only (no sdist on PyPI) and
+# mlx-lm/mlx-vlm publish py3-none-any wheels, so requiring wheels does not break a
+# healthy self-heal; if a wheel is genuinely unavailable the install fails and
+# Studio stays chat-only (the existing safe fallback) until `unsloth studio update`.
+_ONLY_BINARY_ARG = "--only-binary=:all:"
+# Allowlist of environment variables forwarded to the install subprocess. The
+# self-heal runs without confirmation on the default startup path, so it must not
+# hand resolver/build code the full Studio environment. Everything outside this
+# set is dropped, which excludes three dangerous classes by construction:
+#   * secrets (HF_TOKEN, AWS_*, WANDB_API_KEY, ...) that a malicious wheel/sdist
+#     build hook would otherwise read straight out of os.environ;
+#   * package-source redirects (UV_INDEX*, UV_DEFAULT_INDEX, UV_FIND_LINKS,
+#     PIP_INDEX_URL, ...) so a poisoned process env cannot silently repoint the
+#     install at an attacker-controlled index/find-links;
+#   * cache-dir redirects (UV_CACHE_DIR, XDG_CACHE_HOME) so a poisoned env cannot
+#     point uv at an attacker-staged cache (cache poisoning / symlink writes). uv
+#     falls back to its safe user-owned default cache, reused across runs anyway.
+# uv still honours on-disk config (uv.toml / pip.conf), so a corporate mirror
+# configured there keeps working; only process-env redirects are dropped. We set
+# UV_OVERRIDE ourselves in _mlx_install_env, so a poisoned one here is ignored.
+_MLX_ENV_ALLOWLIST = frozenset(
+    {
+        "PATH",
+        "HOME",
+        "USER",
+        "LOGNAME",
+        "TMPDIR",
+        "TMP",
+        "TEMP",
+        "LANG",
+        "LC_ALL",
+        "LC_CTYPE",
+        # proxies + custom CA bundles so installs behind a corporate gateway work
+        "HTTP_PROXY",
+        "HTTPS_PROXY",
+        "NO_PROXY",
+        "ALL_PROXY",
+        "http_proxy",
+        "https_proxy",
+        "no_proxy",
+        "all_proxy",
+        "SSL_CERT_FILE",
+        "SSL_CERT_DIR",
+        "REQUESTS_CA_BUNDLE",
+        "CURL_CA_BUNDLE",
+    }
+)
 _REPAIR_TIMEOUT_S = 900
 
 # Attempt at most once per process; success is sticky (mlx then imports and the
@@ -134,13 +184,22 @@ def _uv_install_cmd(*args: str) -> list[str] | None:
 
 
 def _mlx_install_env() -> dict[str, str]:
-    """Environment for the mlx install. Mirror the main installer
-    (install_python_stack.py) by pointing UV_OVERRIDE at overrides-darwin-arm64.txt,
-    which relaxes mlx-vlm/mlx-lm's transformers>=5 requirement to >=4.57.6. Without
-    it, uv keeps the Studio transformers pin only by silently backtracking mlx-vlm
-    to an old, unsupported version (uv honours UV_OVERRIDE; plain pip ignores it,
-    so the transformers constraint below is the pip-path safety net)."""
-    env = dict(os.environ)
+    """Minimal, allowlisted environment for the unattended mlx install.
+
+    The self-heal runs without confirmation on the default startup path, so it
+    forwards only the variables uv genuinely needs (see _MLX_ENV_ALLOWLIST) instead
+    of the full Studio environment: secrets and package-source redirects in
+    os.environ are dropped so a malicious resolver-selected artifact cannot read
+    Studio secrets or be steered to a hostile index.
+
+    Mirror the main installer (install_python_stack.py) by pointing UV_OVERRIDE at
+    overrides-darwin-arm64.txt, which relaxes mlx-vlm/mlx-lm's transformers>=5
+    requirement to >=4.57.6. Without it, uv keeps the Studio transformers pin only
+    by silently backtracking mlx-vlm to an old, unsupported version (uv honours
+    UV_OVERRIDE; plain pip ignores it, so the transformers constraint below is the
+    pip-path safety net). We set UV_OVERRIDE ourselves, so a poisoned one in the
+    process env is ignored."""
+    env = {key: os.environ[key] for key in _MLX_ENV_ALLOWLIST if key in os.environ}
     override = (
         Path(__file__).resolve().parents[1]
         / "requirements"
@@ -191,7 +250,13 @@ def attempt_mlx_repair(*, timeout: int = _REPAIR_TIMEOUT_S) -> bool:
     constraint_path = None
     try:
         constraint_args, constraint_path = _transformers_constraint_args()
-        cmd = _uv_install_cmd("--upgrade", *_MLX_REINSTALL_ARGS, *constraint_args, *MLX_PACKAGES)
+        cmd = _uv_install_cmd(
+            "--upgrade",
+            _ONLY_BINARY_ARG,
+            *_MLX_REINSTALL_ARGS,
+            *constraint_args,
+            *MLX_PACKAGES,
+        )
         if cmd is None:
             logger.warning(
                 "MLX self-heal requires uv so Studio can apply dependency overrides; "
diff --git a/studio/backend/utils/models/checkpoints.py b/studio/backend/utils/models/checkpoints.py
index 5a992926ec..63f599d0df 100644
--- a/studio/backend/utils/models/checkpoints.py
+++ b/studio/backend/utils/models/checkpoints.py
@@ -4,6 +4,7 @@
 """Checkpoint scanning utilities for discovering training runs and checkpoints."""
 
 import json
+import re
 import structlog
 from loggers import get_logger
 from pathlib import Path
@@ -12,6 +13,22 @@ from utils.paths import outputs_root, resolve_output_dir
 
 logger = get_logger(__name__)
 
+_CHECKPOINT_STEP_RE = re.compile(r"^checkpoint-(\d+)$")
+
+
+def _checkpoint_step(checkpoint_name: str) -> Optional[int]:
+    match = _CHECKPOINT_STEP_RE.fullmatch(checkpoint_name)
+    if match is None:
+        return None
+    return int(match.group(1))
+
+
+def _checkpoint_sort_key(checkpoint_path: Path) -> tuple[int, int, str]:
+    step = _checkpoint_step(checkpoint_path.name)
+    if step is not None:
+        return (0, -step, checkpoint_path.name)
+    return (1, 0, str(checkpoint_path))
+
 
 def _read_checkpoint_loss(checkpoint_path: Path) -> Optional[float]:
     """Read loss from the last log_history entry of trainer_state.json, or None."""
@@ -37,8 +54,10 @@ def scan_checkpoints(
     Returns:
         [(model_name, [(display_name, checkpoint_path, loss), ...], metadata), ...]
         metadata keys (optional): base_model, peft_type, lora_rank.
-        First checkpoint entry is the main adapter; its loss mirrors the last
-        (highest-step) intermediate checkpoint.
+        First checkpoint entry is the main adapter; its loss mirrors the latest
+        (highest-step) intermediate checkpoint. Numbered checkpoints are sorted
+        by numeric step descending; non-numbered checkpoint-* dirs keep the
+        previous lexicographic directory order.
     """
     models = []
     outputs_path = resolve_output_dir(outputs_dir)
@@ -103,18 +122,25 @@ def scan_checkpoints(
             checkpoints.append((item.name, str(item), None))
 
             # Scan for intermediate checkpoints (checkpoint-N subdirs).
-            for sub in sorted(item.iterdir()):
+            valid_checkpoints = []
+            for sub in item.iterdir():
                 if not sub.is_dir() or not sub.name.startswith("checkpoint-"):
                     continue
                 sub_config = sub / "config.json"
                 sub_adapter = sub / "adapter_config.json"
                 if sub_config.exists() or sub_adapter.exists():
-                    loss = _read_checkpoint_loss(sub)
-                    checkpoints.append((sub.name, str(sub), loss))
+                    valid_checkpoints.append(sub)
 
-            # Assign the last checkpoint's loss to the main adapter entry.
-            if len(checkpoints) > 1:
-                last_checkpoint_loss = checkpoints[-1][2]
+            intermediate_checkpoints = []
+            for sub in sorted(valid_checkpoints, key = _checkpoint_sort_key):
+                loss = _read_checkpoint_loss(sub)
+                intermediate_checkpoints.append((sub.name, str(sub), loss))
+
+            checkpoints.extend(intermediate_checkpoints)
+
+            # Assign the latest checkpoint's loss to the main adapter entry.
+            if intermediate_checkpoints:
+                last_checkpoint_loss = intermediate_checkpoints[0][2]
                 checkpoints[0] = (
                     checkpoints[0][0],
                     checkpoints[0][1],
diff --git a/studio/frontend/package-lock.json b/studio/frontend/package-lock.json
index a521552d79..6202d3ce22 100644
--- a/studio/frontend/package-lock.json
+++ b/studio/frontend/package-lock.json
@@ -88,7 +88,7 @@
         "globals": "^17.4.0",
         "typescript": "~5.9.3",
         "typescript-eslint": "^8.55.0",
-        "vite": "^8.0.1"
+        "vite": "^8.0.16"
       },
       "engines": {
         "node": "^20.19.0 || >=22.12.0"
@@ -1913,13 +1913,13 @@
       }
     },
     "node_modules/@napi-rs/wasm-runtime": {
-      "version": "1.1.4",
-      "resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.4.tgz",
-      "integrity": "sha512-3NQNNgA1YSlJb/kMH1ildASP9HW7/7kYnRI2szWJaofaS1hWmbGI4H+d3+22aGzXXN9IJ+n+GiFVcGipJP18ow==",
+      "version": "1.1.5",
+      "resolved": "https://registry.npmjs.org/@napi-rs/wasm-runtime/-/wasm-runtime-1.1.5.tgz",
+      "integrity": "sha512-AWPoBRJ9tsnVhor4sjO7rkni+7p+2IAEFj6cx06UgP10jkQHqay/36uRV/bFkgrh18D9vb4cr8Q0Pthskgzy+Q==",
       "license": "MIT",
       "optional": true,
       "dependencies": {
-        "@tybys/wasm-util": "^0.10.1"
+        "@tybys/wasm-util": "^0.10.2"
       },
       "funding": {
         "type": "github",
@@ -2027,9 +2027,9 @@
       "license": "MIT"
     },
     "node_modules/@oxc-project/types": {
-      "version": "0.127.0",
-      "resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.127.0.tgz",
-      "integrity": "sha512-aIYXQBo4lCbO4z0R3FHeucQHpF46l2LbMdxRvqvuRuW2OxdnSkcng5B8+K12spgLDj93rtN3+J2Vac/TIO+ciQ==",
+      "version": "0.133.0",
+      "resolved": "https://registry.npmjs.org/@oxc-project/types/-/types-0.133.0.tgz",
+      "integrity": "sha512-KzkdCd6Uxqnf6l3HOw1xfatAlUURA0g14cvBYFyJ5SaNOQbOUvBr9PKArcPcrNIeRsBdgcUzOGrhKveVpvOIGA==",
       "license": "MIT",
       "funding": {
         "url": "https://github.com/sponsors/Boshen"
@@ -5583,9 +5583,9 @@
       }
     },
     "node_modules/@rolldown/binding-android-arm64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-s70pVGhw4zqGeFnXWvAzJDlvxhlRollagdCCKRgOsgUOH3N1l0LIxf83AtGzmb5SiVM4Hjl5HyarMRfdfj3DaQ==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-android-arm64/-/binding-android-arm64-1.0.3.tgz",
+      "integrity": "sha512-454rs7jHngixp/NMxd5srYD57OnzSlZ/eFTETjORQHLwJG1lRtmNOJcBerZlfu4GjKqeq8aCCIQrMdHyhI51Hw==",
       "cpu": [
         "arm64"
       ],
@@ -5599,9 +5599,9 @@
       }
     },
     "node_modules/@rolldown/binding-darwin-arm64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-4ksWc9n0mhlZpZ9PMZgTGjeOPRu8MB1Z3Tz0Mo02eWfWCHMW1zN82Qz/pL/rC+yQa+8ZnutMF0JjJe7PjwasYw==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-arm64/-/binding-darwin-arm64-1.0.3.tgz",
+      "integrity": "sha512-PcAhP+ynjURNyy8SKGl5DQP94aGuB/7JrXJb/t7P+hanXvQVMWzUvRRhBAcg/lNRadBhoUPqSoP4xw5tR/KBEA==",
       "cpu": [
         "arm64"
       ],
@@ -5615,9 +5615,9 @@
       }
     },
     "node_modules/@rolldown/binding-darwin-x64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-SUSDOI6WwUVNcWxd02QEBjLdY1VPHvlEkw6T/8nYG322iYWCTxRb1vzk4E+mWWYehTp7ERibq54LSJGjmouOsw==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-darwin-x64/-/binding-darwin-x64-1.0.3.tgz",
+      "integrity": "sha512-9YpfeUvSE2RS7wysJ81uOZkXJz7f7Q55H2Gvp3VEw/EsahqDtrphrZ0EwDLK5vvKOzaCrBsjF8JmnMLcUt78Gg==",
       "cpu": [
         "x64"
       ],
@@ -5631,9 +5631,9 @@
       }
     },
     "node_modules/@rolldown/binding-freebsd-x64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-hwnz3nw9dbJ05EDO/PvcjaaewqqDy7Y1rn1UO81l8iIK1GjenME75dl16ajbvSSMfv66WXSRCYKIqfgq2KCfxw==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-freebsd-x64/-/binding-freebsd-x64-1.0.3.tgz",
+      "integrity": "sha512-yB1IlAsSNHncV6SCTL27/MVGR5htvQsoGxIv5KMGXALp+Ll1wYsn+x98M9MW7qa+NdSbvrrY7ANI4wLJ0n1e6g==",
       "cpu": [
         "x64"
       ],
@@ -5647,9 +5647,9 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm-gnueabihf": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.0.0-rc.17.tgz",
-      "integrity": "sha512-IS+W7epTcwANmFSQFrS1SivEXHtl1JtuQA9wlxrZTcNi6mx+FDOYrakGevvvTwgj2JvWiK8B29/qD9BELZPyXQ==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm-gnueabihf/-/binding-linux-arm-gnueabihf-1.0.3.tgz",
+      "integrity": "sha512-Yi30IVAAfLUCy2MseFjbB1jAMDl1VMCAas5StnYp8da9+CKvMd2H2cbEjWcw5NPaPqzvYkVIaF1nNUG+b7u/sw==",
       "cpu": [
         "arm"
       ],
@@ -5663,12 +5663,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm64-gnu": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.0.0-rc.17.tgz",
-      "integrity": "sha512-e6usGaHKW5BMNZOymS1UcEYGowQMWcgZ71Z17Sl/h2+ZziNJ1a9n3Zvcz6LdRyIW5572wBCTH/Z+bKuZouGk9Q==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-gnu/-/binding-linux-arm64-gnu-1.0.3.tgz",
+      "integrity": "sha512-jsO7R8To+AdlYgUmN5sHSCZbfhtMBkO0WUx8iORQnPcMMdgr7qM2DQmMwgabs3GhNztdmoKkMKQFHD6DTMCIQw==",
       "cpu": [
         "arm64"
       ],
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5679,12 +5682,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-arm64-musl": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.0.0-rc.17.tgz",
-      "integrity": "sha512-b/CgbwAJpmrRLp02RPfhbudf5tZnN9nsPWK82znefso832etkem8H7FSZwxrOI9djcdTP7U6YfNhbRnh7djErg==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-arm64-musl/-/binding-linux-arm64-musl-1.0.3.tgz",
+      "integrity": "sha512-VWkUHwWriDciit80wleYwKILoR/KMvxh/IdwS/paX+ZgpuRpCrKLUdadJbc0NpBEiyhpYawsJ73j9aCvOH+f7Q==",
       "cpu": [
         "arm64"
       ],
+      "libc": [
+        "musl"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5695,12 +5701,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-ppc64-gnu": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.0.0-rc.17.tgz",
-      "integrity": "sha512-4EII1iNGRUN5WwGbF/kOh/EIkoDN9HsupgLQoXfY+D1oyJm7/F4t5PYU5n8SWZgG0FEwakyM8pGgwcBYruGTlA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-ppc64-gnu/-/binding-linux-ppc64-gnu-1.0.3.tgz",
+      "integrity": "sha512-5f1laC0SlIR0yDbFCd8acUhvJIag6N3zC5P7oUPN6wX0aOma+uKJ0wBDH5aq7I1PVI2ttTlhJwzwRIBnLiSGEg==",
       "cpu": [
         "ppc64"
       ],
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5711,12 +5720,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-s390x-gnu": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.0.0-rc.17.tgz",
-      "integrity": "sha512-AH8oq3XqQo4IibpVXvPeLDI5pzkpYn0WiZAfT05kFzoJ6tQNzwRdDYQ45M8I/gslbodRZwW8uxLhbSBbkv96rA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-s390x-gnu/-/binding-linux-s390x-gnu-1.0.3.tgz",
+      "integrity": "sha512-Iq4ko0r4XsgbrF/LunNgHtAGLRRVE2kXonAXQ/MV0mC6jQpMOhW1SvtZja2EhC/kd05++bP78dsqBeIQyYJ6Yg==",
       "cpu": [
         "s390x"
       ],
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5727,12 +5739,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-x64-gnu": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.0.0-rc.17.tgz",
-      "integrity": "sha512-cLnjV3xfo7KslbU41Z7z8BH/E1y5mzUYzAqih1d1MDaIGZRCMqTijqLv76/P7fyHuvUcfGsIpqCdddbxLLK9rA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-gnu/-/binding-linux-x64-gnu-1.0.3.tgz",
+      "integrity": "sha512-B8m6tD5+/N5FeNQFbKlLA/2yVq9ycQP1SeedyEYYKWBNR3ZQbkvIUcNnDNM03lO1l5F2roiiFJGgvoLLyZXtSg==",
       "cpu": [
         "x64"
       ],
+      "libc": [
+        "glibc"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5743,12 +5758,15 @@
       }
     },
     "node_modules/@rolldown/binding-linux-x64-musl": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.0.0-rc.17.tgz",
-      "integrity": "sha512-0phclDw1spsL7dUB37sIARuis2tAgomCJXAHZlpt8PXZ4Ba0dRP1e+66lsRqrfhISeN9bEGNjQs+T/Fbd7oYGw==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-linux-x64-musl/-/binding-linux-x64-musl-1.0.3.tgz",
+      "integrity": "sha512-pSdpdUJHkuCxun9LE7jvgUB9qsRgaiyNNCX7m/AvHTcq67AiT/Yhoxvw5zPfhrM8k/BfP8ce/hMOpthKDpEUow==",
       "cpu": [
         "x64"
       ],
+      "libc": [
+        "musl"
+      ],
       "license": "MIT",
       "optional": true,
       "os": [
@@ -5759,9 +5777,9 @@
       }
     },
     "node_modules/@rolldown/binding-openharmony-arm64": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.0.0-rc.17.tgz",
-      "integrity": "sha512-0ag/hEgXOwgw4t8QyQvUCxvEg+V0KBcA6YuOx9g0r02MprutRF5dyljgm3EmR02O292UX7UeS6HzWHAl6KgyhA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-openharmony-arm64/-/binding-openharmony-arm64-1.0.3.tgz",
+      "integrity": "sha512-OXXS3RKJgX2uLwM+gYyuH5omcH8fL1LJs96pZGgtetVCahON57+d4SJHzTgZiOjxgGkSnpXpOsWuPDGAKAigEg==",
       "cpu": [
         "arm64"
       ],
@@ -5775,9 +5793,9 @@
       }
     },
     "node_modules/@rolldown/binding-wasm32-wasi": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-wasm32-wasi/-/binding-wasm32-wasi-1.0.0-rc.17.tgz",
-      "integrity": "sha512-LEXei6vo0E5wTGwpkJ4KoT3OZJRnglwldt5ziLzOlc6qqb55z4tWNq2A+PFqCJuvWWdP53CVhG1Z9NtToDPJrA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-wasm32-wasi/-/binding-wasm32-wasi-1.0.3.tgz",
+      "integrity": "sha512-JTtb8BWFynicNSoPrehsCzBtOKjZ6jhMiPFEmOiuXg1Fl8dn2KHQob+GuPSGR0dryQa1PQJbzjF3dqO/whhjLg==",
       "cpu": [
         "wasm32"
       ],
@@ -5793,9 +5811,9 @@
       }
     },
     "node_modules/@rolldown/binding-win32-arm64-msvc": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.0.0-rc.17.tgz",
-      "integrity": "sha512-gUmyzBl3SPMa6hrqFUth9sVfcLBlYsbMzBx5PlexMroZStgzGqlZ26pYG89rBb45Mnia+oil6YAIFeEWGWhoZA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-arm64-msvc/-/binding-win32-arm64-msvc-1.0.3.tgz",
+      "integrity": "sha512-gEdFFEN70A/jxb2svrWsN3aDL7OUtmvlOy+6fa2jxG8K0wQ1ZbdeLGnidov6Yu5/733dI5ySfzFlQ/cb0bSz1g==",
       "cpu": [
         "arm64"
       ],
@@ -5809,9 +5827,9 @@
       }
     },
     "node_modules/@rolldown/binding-win32-x64-msvc": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.0.0-rc.17.tgz",
-      "integrity": "sha512-3hkiolcUAvPB9FLb3UZdfjVVNWherN1f/skkGWJP/fgSQhYUZpSIRr0/I8ZK9TkF3F7kxvJAk0+IcKvPHk9qQg==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/@rolldown/binding-win32-x64-msvc/-/binding-win32-x64-msvc-1.0.3.tgz",
+      "integrity": "sha512-eXB7CHuaQdqmJcc3koCNtNPmT/bj2gc999kUFgBxG8Ac0NdgXc4rkCHhqrgrhN3zddvvvrgzj1e90SuSfmyIXA==",
       "cpu": [
         "x64"
       ],
@@ -13091,6 +13109,34 @@
         "points-on-curve": "0.2.0"
       }
     },
+    "node_modules/postcss": {
+      "version": "8.5.15",
+      "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.15.tgz",
+      "integrity": "sha512-FfR8sjd4em2T6fb3I2MwAJU7HWVMr9zba+enmQeeWFfCbm+UOC/0X4DS8XtpUTMwWMGbjKYP7xjfNekzyGmB3A==",
+      "funding": [
+        {
+          "type": "opencollective",
+          "url": "https://opencollective.com/postcss/"
+        },
+        {
+          "type": "tidelift",
+          "url": "https://tidelift.com/funding/github/npm/postcss"
+        },
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/ai"
+        }
+      ],
+      "license": "MIT",
+      "dependencies": {
+        "nanoid": "^3.3.12",
+        "picocolors": "^1.1.1",
+        "source-map-js": "^1.2.1"
+      },
+      "engines": {
+        "node": "^10 || ^12 || >=14"
+      }
+    },
     "node_modules/postcss-selector-parser": {
       "version": "7.1.1",
       "resolved": "https://registry.npmjs.org/postcss-selector-parser/-/postcss-selector-parser-7.1.1.tgz",
@@ -13104,6 +13150,24 @@
         "node": ">=4"
       }
     },
+    "node_modules/postcss/node_modules/nanoid": {
+      "version": "3.3.12",
+      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.12.tgz",
+      "integrity": "sha512-ZB9RH/39qpq5Vu6Y+NmUaFhQR6pp+M2Xt76XBnEwDaGcVAqhlvxrl3B2bKS5D3NH3QR76v3aSrKaF/Kiy7lEtQ==",
+      "funding": [
+        {
+          "type": "github",
+          "url": "https://github.com/sponsors/ai"
+        }
+      ],
+      "license": "MIT",
+      "bin": {
+        "nanoid": "bin/nanoid.cjs"
+      },
+      "engines": {
+        "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
+      }
+    },
     "node_modules/powershell-utils": {
       "version": "0.1.0",
       "resolved": "https://registry.npmjs.org/powershell-utils/-/powershell-utils-0.1.0.tgz",
@@ -13977,13 +14041,13 @@
       "license": "Unlicense"
     },
     "node_modules/rolldown": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.0.0-rc.17.tgz",
-      "integrity": "sha512-ZrT53oAKrtA4+YtBWPQbtPOxIbVDbxT0orcYERKd63VJTF13zPcgXTvD4843L8pcsI7M6MErt8QtON6lrB9tyA==",
+      "version": "1.0.3",
+      "resolved": "https://registry.npmjs.org/rolldown/-/rolldown-1.0.3.tgz",
+      "integrity": "sha512-i00lAJ2ks1BYr7rjNjKC7BcqAS7nVfiT3QX1SI5aY+AFHblCmaUf9OE9dbdzDvW6dJxbi2ZCZiy9v3CcwOiX3g==",
       "license": "MIT",
       "dependencies": {
-        "@oxc-project/types": "=0.127.0",
-        "@rolldown/pluginutils": "1.0.0-rc.17"
+        "@oxc-project/types": "=0.133.0",
+        "@rolldown/pluginutils": "^1.0.0"
       },
       "bin": {
         "rolldown": "bin/cli.mjs"
@@ -13992,27 +14056,27 @@
         "node": "^20.19.0 || >=22.12.0"
       },
       "optionalDependencies": {
-        "@rolldown/binding-android-arm64": "1.0.0-rc.17",
-        "@rolldown/binding-darwin-arm64": "1.0.0-rc.17",
-        "@rolldown/binding-darwin-x64": "1.0.0-rc.17",
-        "@rolldown/binding-freebsd-x64": "1.0.0-rc.17",
-        "@rolldown/binding-linux-arm-gnueabihf": "1.0.0-rc.17",
-        "@rolldown/binding-linux-arm64-gnu": "1.0.0-rc.17",
-        "@rolldown/binding-linux-arm64-musl": "1.0.0-rc.17",
-        "@rolldown/binding-linux-ppc64-gnu": "1.0.0-rc.17",
-        "@rolldown/binding-linux-s390x-gnu": "1.0.0-rc.17",
-        "@rolldown/binding-linux-x64-gnu": "1.0.0-rc.17",
-        "@rolldown/binding-linux-x64-musl": "1.0.0-rc.17",
-        "@rolldown/binding-openharmony-arm64": "1.0.0-rc.17",
-        "@rolldown/binding-wasm32-wasi": "1.0.0-rc.17",
-        "@rolldown/binding-win32-arm64-msvc": "1.0.0-rc.17",
-        "@rolldown/binding-win32-x64-msvc": "1.0.0-rc.17"
+        "@rolldown/binding-android-arm64": "1.0.3",
+        "@rolldown/binding-darwin-arm64": "1.0.3",
+        "@rolldown/binding-darwin-x64": "1.0.3",
+        "@rolldown/binding-freebsd-x64": "1.0.3",
+        "@rolldown/binding-linux-arm-gnueabihf": "1.0.3",
+        "@rolldown/binding-linux-arm64-gnu": "1.0.3",
+        "@rolldown/binding-linux-arm64-musl": "1.0.3",
+        "@rolldown/binding-linux-ppc64-gnu": "1.0.3",
+        "@rolldown/binding-linux-s390x-gnu": "1.0.3",
+        "@rolldown/binding-linux-x64-gnu": "1.0.3",
+        "@rolldown/binding-linux-x64-musl": "1.0.3",
+        "@rolldown/binding-openharmony-arm64": "1.0.3",
+        "@rolldown/binding-wasm32-wasi": "1.0.3",
+        "@rolldown/binding-win32-arm64-msvc": "1.0.3",
+        "@rolldown/binding-win32-x64-msvc": "1.0.3"
       }
     },
     "node_modules/rolldown/node_modules/@rolldown/pluginutils": {
-      "version": "1.0.0-rc.17",
-      "resolved": "https://registry.npmjs.org/@rolldown/pluginutils/-/pluginutils-1.0.0-rc.17.tgz",
-      "integrity": "sha512-n8iosDOt6Ig1UhJ2AYqoIhHWh/isz0xpicHTzpKBeotdVsTEcxsSA/i3EVM7gQAj0rU27OLAxCjzlj15IWY7bg==",
+      "version": "1.0.1",
+      "resolved": "https://registry.npmjs.org/@rolldown/pluginutils/-/pluginutils-1.0.1.tgz",
+      "integrity": "sha512-2j9bGt5Jh8hj+vPtgzPtl72j0yRxHAyumoo6TNfAjsLB04UtpSvPbPcDcBMxz7n+9CYB0c1GxQFxYRg2jimqGw==",
       "license": "MIT"
     },
     "node_modules/roughjs": {
@@ -14275,52 +14339,6 @@
         "node": ">=20"
       }
     },
-    "node_modules/shadcn/node_modules/nanoid": {
-      "version": "3.3.12",
-      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.12.tgz",
-      "integrity": "sha512-ZB9RH/39qpq5Vu6Y+NmUaFhQR6pp+M2Xt76XBnEwDaGcVAqhlvxrl3B2bKS5D3NH3QR76v3aSrKaF/Kiy7lEtQ==",
-      "funding": [
-        {
-          "type": "github",
-          "url": "https://github.com/sponsors/ai"
-        }
-      ],
-      "license": "MIT",
-      "bin": {
-        "nanoid": "bin/nanoid.cjs"
-      },
-      "engines": {
-        "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
-      }
-    },
-    "node_modules/shadcn/node_modules/postcss": {
-      "version": "8.5.14",
-      "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.14.tgz",
-      "integrity": "sha512-SoSL4+OSEtR99LHFZQiJLkT59C5B1amGO1NzTwj7TT1qCUgUO6hxOvzkOYxD+vMrXBM3XJIKzokoERdqQq/Zmg==",
-      "funding": [
-        {
-          "type": "opencollective",
-          "url": "https://opencollective.com/postcss/"
-        },
-        {
-          "type": "tidelift",
-          "url": "https://tidelift.com/funding/github/npm/postcss"
-        },
-        {
-          "type": "github",
-          "url": "https://github.com/sponsors/ai"
-        }
-      ],
-      "license": "MIT",
-      "dependencies": {
-        "nanoid": "^3.3.11",
-        "picocolors": "^1.1.1",
-        "source-map-js": "^1.2.1"
-      },
-      "engines": {
-        "node": "^10 || ^12 || >=14"
-      }
-    },
     "node_modules/shadcn/node_modules/zod": {
       "version": "3.25.76",
       "resolved": "https://registry.npmjs.org/zod/-/zod-3.25.76.tgz",
@@ -14786,9 +14804,9 @@
       }
     },
     "node_modules/tinyglobby": {
-      "version": "0.2.16",
-      "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.16.tgz",
-      "integrity": "sha512-pn99VhoACYR8nFHhxqix+uvsbXineAasWm5ojXoN8xEwK5Kd3/TrhNn1wByuD52UxWRLy8pu+kRMniEi6Eq9Zg==",
+      "version": "0.2.17",
+      "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.17.tgz",
+      "integrity": "sha512-wXR/dYpcqKmfWpEdZjiKJOwCNFndD0DMnrW/cYjVGttEkBfVgcLFHoNrlj47mjOVic9yyNu65alsgF4NQyTa2g==",
       "license": "MIT",
       "dependencies": {
         "fdir": "^6.5.0",
@@ -15456,16 +15474,16 @@
       }
     },
     "node_modules/vite": {
-      "version": "8.0.10",
-      "resolved": "https://registry.npmjs.org/vite/-/vite-8.0.10.tgz",
-      "integrity": "sha512-rZuUu9j6J5uotLDs+cAA4O5H4K1SfPliUlQwqa6YEwSrWDZzP4rhm00oJR5snMewjxF5V/K3D4kctsUTsIU9Mw==",
+      "version": "8.0.16",
+      "resolved": "https://registry.npmjs.org/vite/-/vite-8.0.16.tgz",
+      "integrity": "sha512-h9bXPmJichP5fLmVQo3PyaGSDE2n3aPuomeAlVRm0JLmt4rY6zmPKd59HYI4LNW8oTK7tlTsuC7l/m7awx9Jcw==",
       "license": "MIT",
       "dependencies": {
         "lightningcss": "^1.32.0",
         "picomatch": "^4.0.4",
-        "postcss": "^8.5.10",
-        "rolldown": "1.0.0-rc.17",
-        "tinyglobby": "^0.2.16"
+        "postcss": "^8.5.15",
+        "rolldown": "1.0.3",
+        "tinyglobby": "^0.2.17"
       },
       "bin": {
         "vite": "bin/vite.js"
@@ -15481,7 +15499,7 @@
       },
       "peerDependencies": {
         "@types/node": "^20.19.0 || >=22.12.0",
-        "@vitejs/devtools": "^0.1.0",
+        "@vitejs/devtools": "^0.1.18",
         "esbuild": "^0.27.0 || ^0.28.0",
         "jiti": ">=1.21.0",
         "less": "^4.0.0",
@@ -15546,52 +15564,6 @@
         "node": "^8.16.0 || ^10.6.0 || >=11.0.0"
       }
     },
-    "node_modules/vite/node_modules/nanoid": {
-      "version": "3.3.12",
-      "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.12.tgz",
-      "integrity": "sha512-ZB9RH/39qpq5Vu6Y+NmUaFhQR6pp+M2Xt76XBnEwDaGcVAqhlvxrl3B2bKS5D3NH3QR76v3aSrKaF/Kiy7lEtQ==",
-      "funding": [
-        {
-          "type": "github",
-          "url": "https://github.com/sponsors/ai"
-        }
-      ],
-      "license": "MIT",
-      "bin": {
-        "nanoid": "bin/nanoid.cjs"
-      },
-      "engines": {
-        "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
-      }
-    },
-    "node_modules/vite/node_modules/postcss": {
-      "version": "8.5.14",
-      "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.14.tgz",
-      "integrity": "sha512-SoSL4+OSEtR99LHFZQiJLkT59C5B1amGO1NzTwj7TT1qCUgUO6hxOvzkOYxD+vMrXBM3XJIKzokoERdqQq/Zmg==",
-      "funding": [
-        {
-          "type": "opencollective",
-          "url": "https://opencollective.com/postcss/"
-        },
-        {
-          "type": "tidelift",
-          "url": "https://tidelift.com/funding/github/npm/postcss"
-        },
-        {
-          "type": "github",
-          "url": "https://github.com/sponsors/ai"
-        }
-      ],
-      "license": "MIT",
-      "dependencies": {
-        "nanoid": "^3.3.11",
-        "picocolors": "^1.1.1",
-        "source-map-js": "^1.2.1"
-      },
-      "engines": {
-        "node": "^10 || ^12 || >=14"
-      }
-    },
     "node_modules/warning": {
       "version": "4.0.3",
       "resolved": "https://registry.npmjs.org/warning/-/warning-4.0.3.tgz",
diff --git a/studio/frontend/package.json b/studio/frontend/package.json
index c49b62ab50..0956c710f5 100644
--- a/studio/frontend/package.json
+++ b/studio/frontend/package.json
@@ -107,7 +107,7 @@
     "globals": "^17.4.0",
     "typescript": "~5.9.3",
     "typescript-eslint": "^8.55.0",
-    "vite": "^8.0.1"
+    "vite": "^8.0.16"
   },
   "allowScripts": {
     "@biomejs/biome@1.9.4": true,
diff --git a/studio/frontend/src/app/routes/__root.tsx b/studio/frontend/src/app/routes/__root.tsx
index b7e7bc01d2..77ba5788db 100644
--- a/studio/frontend/src/app/routes/__root.tsx
+++ b/studio/frontend/src/app/routes/__root.tsx
@@ -182,7 +182,9 @@ function RootLayout() {
         chatRuntime.setActiveThreadId(null);
         chatRuntime.setActiveProjectId(null);
         chatRuntime.setIncognito(false);
-        if (chatRuntime.pendingSelection) chatRuntime.abandonStagedModel();
+        // Detach the staging UI but keep any in-flight download running, like Hub.
+        if (chatRuntime.pendingSelection)
+          chatRuntime.abandonStagedModel({ keepDownload: true });
         void navigate({
           to: "/chat",
           search: { new: crypto.randomUUID() },
@@ -205,7 +207,10 @@ function RootLayout() {
     chatRuntime.setActiveProjectId(null);
     chatRuntime.setActiveThreadId(null);
     chatRuntime.setIncognito(false);
-    if (chatRuntime.pendingSelection) chatRuntime.abandonStagedModel();
+    // Leaving chat must not kill an in-flight download: detach the staging UI
+    // but keep the transfer running in the manager, like a Hub download.
+    if (chatRuntime.pendingSelection)
+      chatRuntime.abandonStagedModel({ keepDownload: true });
   }, [isChatRoute]);
 
   return (
diff --git a/studio/frontend/src/components/app-sidebar.tsx b/studio/frontend/src/components/app-sidebar.tsx
index 734860138a..71956ff9db 100644
--- a/studio/frontend/src/components/app-sidebar.tsx
+++ b/studio/frontend/src/components/app-sidebar.tsx
@@ -45,8 +45,11 @@ import { Spinner } from "@/components/ui/spinner";
 import { Switch } from "@/components/ui/switch";
 import { useAnimatedThemeToggle } from "@/components/ui/animated-theme-toggler";
 import { cn } from "@/lib/utils";
+import { useWebUpdateCheck } from "@/hooks/use-web-update-check";
 import {
   Archive03Icon,
+  ArrowRight02Icon,
+  BadgeInfoIcon,
   ChefHatIcon,
   CursorInfo02Icon,
   DashboardCircleIcon,
@@ -253,6 +256,19 @@ function NavItem({
   );
 }
 
+// TEMP DEV override: preview the update card on installs with no real update
+// (e.g. an editable checkout). In the browser console run
+// `localStorage.setItem("unsloth_force_update_card", "1")` and reload. Remove
+// before merge.
+function devForceUpdateCard(): boolean {
+  if (typeof window === "undefined") return false;
+  try {
+    return window.localStorage.getItem("unsloth_force_update_card") === "1";
+  } catch {
+    return false;
+  }
+}
+
 export function AppSidebar() {
   const t = useT();
   const { isDark, toggleTheme, anchorRef } = useAnimatedThemeToggle();
@@ -265,6 +281,16 @@ export function AppSidebar() {
   const { togglePinned, isMobile, setOpenMobile } = useSidebar();
   const navigate = useNavigate();
 
+  // Web update detection: `webUpdate` is non-null only when the installed
+  // (PyPI) version is behind the latest release, so the card is hidden by
+  // default. `forceUpdateCard` is a TEMP dev override to preview it on installs
+  // with no real update (e.g. an editable checkout); remove before merge.
+  const { status: webUpdate } = useWebUpdateCheck();
+  const [forceUpdateCard] = useState(devForceUpdateCard);
+  const showUpdateCard = Boolean(webUpdate) || forceUpdateCard;
+  const updateVersion =
+    webUpdate?.latestVersion ?? (forceUpdateCard ? "0.0.0" : null);
+
   // Auto-close mobile Sheet after navigation
   const closeMobileIfOpen = () => {
     if (isMobile) setOpenMobile(false);
@@ -1348,7 +1374,7 @@ export function AppSidebar() {
         )}
       </SidebarContent>
 
-      <SidebarFooter className="relative group-data-[collapsible=icon]:px-0">
+      <SidebarFooter className="relative pt-3 pb-4 group-data-[collapsible=icon]:px-0">
         {/* Fade above the profile box, shown only when there's more list below
             the fold; at the bottom (or short lists) it fades so the last row
             shows fully (Gemini-style). right-2 keeps it clear of the 8px scrollbar gutter. */}
@@ -1359,7 +1385,54 @@ export function AppSidebar() {
             canScrollDown ? "opacity-100" : "opacity-0",
           )}
         />
-        <SidebarMenu>
+        <SidebarMenu className="gap-3 group-data-[collapsible=icon]:gap-2.5">
+          {/* Update affordance — shows only when a newer version is available. */}
+          {showUpdateCard && (
+            <SidebarMenuItem>
+              <button
+                type="button"
+                aria-label={t("shell.updateAvailable")}
+                onClick={() => {
+                  useSettingsDialogStore
+                    .getState()
+                    .openDialog("about", { scrollTarget: "about-updates" });
+                  closeMobileIfOpen();
+                }}
+                className="flex h-[44px] w-full items-center gap-[9px] rounded-[14px] border border-border/60 bg-transparent px-2 py-[3px] text-left transition-colors hover:bg-nav-surface-hover focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring group-data-[collapsible=icon]:mx-auto group-data-[collapsible=icon]:h-[34px] group-data-[collapsible=icon]:w-[34px] group-data-[collapsible=icon]:justify-center group-data-[collapsible=icon]:gap-0 group-data-[collapsible=icon]:rounded-full group-data-[collapsible=icon]:p-0"
+              >
+                <span
+                  aria-hidden="true"
+                  className="flex size-[32px] shrink-0 items-center justify-center group-data-[collapsible=icon]:size-full"
+                >
+                  <HugeiconsIcon
+                    icon={BadgeInfoIcon}
+                    strokeWidth={1.75}
+                    className="size-[21px] text-nav-fg"
+                  />
+                </span>
+                <div className="flex min-w-0 flex-col gap-px leading-tight group-data-[collapsible=icon]:hidden">
+                  <span className="truncate font-heading text-[13.5px] font-semibold text-nav-fg">
+                    {t("shell.updateAvailable")}
+                  </span>
+                  {updateVersion && (
+                    <span className="truncate text-[11.5px] text-muted-foreground">
+                      v{updateVersion}
+                    </span>
+                  )}
+                </div>
+                <span
+                  aria-hidden="true"
+                  className="ml-auto flex size-[32px] shrink-0 items-center justify-center text-muted-foreground group-data-[collapsible=icon]:hidden"
+                >
+                  <HugeiconsIcon
+                    icon={ArrowRight02Icon}
+                    className="size-[17px]"
+                    strokeWidth={1.75}
+                  />
+                </span>
+              </button>
+            </SidebarMenuItem>
+          )}
           <SidebarMenuItem>
             <DropdownMenu>
               <DropdownMenuTrigger asChild>
@@ -1381,11 +1454,16 @@ export function AppSidebar() {
                     <span className="truncate text-[11.5px] tracking-nav text-muted-foreground">Unsloth</span>
                   </div>
                   {/* settings cog (replaces the up/down chevron) */}
-                  <HugeiconsIcon
-                    icon={Settings02Icon}
-                    strokeWidth={1.5}
-                    className="ml-auto !size-[18px] text-muted-foreground group-data-[collapsible=icon]:hidden"
-                  />
+                  <span
+                    aria-hidden="true"
+                    className="ml-auto flex size-[32px] shrink-0 items-center justify-center text-muted-foreground group-data-[collapsible=icon]:hidden"
+                  >
+                    <HugeiconsIcon
+                      icon={Settings02Icon}
+                      strokeWidth={1.5}
+                      className="!size-[18px]"
+                    />
+                  </span>
                 </SidebarMenuButton>
               </DropdownMenuTrigger>
               <DropdownMenuContent
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
index 46b6fe2e63..90c1c04106 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
+++ b/studio/frontend/src/components/assistant-ui/model-selector/pickers.tsx
@@ -2336,7 +2336,7 @@ export function HubModelPicker({
           <Input
             value={query}
             onChange={(event) => setQuery(event.target.value)}
-            placeholder="Search models"
+            placeholder="Search Unsloth models"
             data-model-picker-search-input={true}
             className="field-soft h-9 border-0 pl-8 pr-8"
           />
@@ -2345,15 +2345,20 @@ export function HubModelPicker({
           )}
         </div>
         {onBrowseHub ? (
-          <button
-            type="button"
-            onClick={onBrowseHub}
-            aria-label="Search more models on the Hub"
-            className="hub-tab-toggle-pill flex h-9 w-[110px] shrink-0 items-center justify-center gap-[5px] rounded-full border-0 text-xs text-foreground transition-colors"
-          >
-            <HugeiconsIcon icon={DashboardCircleIcon} className="size-4" />
-            Search Hub
-          </button>
+          <Tooltip>
+            <TooltipTrigger asChild>
+              <button
+                type="button"
+                onClick={onBrowseHub}
+                aria-label="Search more models on the Hub"
+                className="hub-tab-toggle-pill flex h-9 w-[110px] shrink-0 items-center justify-center gap-[5px] rounded-full border-0 text-xs text-foreground transition-colors"
+              >
+                <HugeiconsIcon icon={DashboardCircleIcon} className="size-4" />
+                Search Hub
+              </button>
+            </TooltipTrigger>
+            <TooltipContent>Search all models</TooltipContent>
+          </Tooltip>
         ) : null}
       </div>
 
@@ -2386,7 +2391,7 @@ export function HubModelPicker({
           // Height tracks the content up to the cap, so short lists do not
           // leave white space. scroll-py + symmetric px keep the focus ring off
           // the overflow clip edges during keyboard nav.
-          "model-list-scroll max-h-[21rem] overflow-y-auto scroll-py-1.5 px-0.5 mr-1",
+          "model-list-scroll max-h-[335px] overflow-y-auto scroll-py-1.5 px-0.5 mr-1",
           listScrolled && "is-scrolled",
           listMoreBelow && "is-bottom-faded",
         )}
@@ -3362,7 +3367,7 @@ export function HubModelPicker({
       {/* Floating eject pill: overlaid on the list bottom, outside the scroll
           so the edge fade never touches it. Only the pill catches clicks. */}
       {onEject ? (
-        <div className="pointer-events-none absolute inset-x-0 bottom-0 flex justify-end pr-3.5 pb-5">
+        <div className="pointer-events-none absolute inset-x-0 bottom-0 flex justify-end pr-3.5 pb-[19px]">
           <button
             type="button"
             onClick={onEject}
diff --git a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
index 85e088ffb2..ec75b17f20 100644
--- a/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
+++ b/studio/frontend/src/components/assistant-ui/model-selector/remembered-load-settings.ts
@@ -14,6 +14,21 @@ export interface RememberedLoadSettings {
   tensorParallel: boolean;
 }
 
+// Storage key for a pick's remembered settings. The remembered knobs are
+// VRAM-budget driven (context override, KV-cache dtype, tensor-parallel), so the
+// right values differ per quant. An HF repo collapses all its GGUF variants into
+// one `id`, so fold the variant in to scope settings per quant. Local .gguf
+// paths key by their file path (already file-specific); native drag-drop files
+// key by display label, so same-named files in different folders share an entry.
+export function rememberedLoadSettingsKey(selection: {
+  id: string;
+  ggufVariant?: string | null;
+}): string {
+  return selection.ggufVariant
+    ? `${selection.id}::${selection.ggufVariant}`
+    : selection.id;
+}
+
 function readAll(): Record<string, RememberedLoadSettings> {
   try {
     return JSON.parse(localStorage.getItem(KEY) ?? "{}");
@@ -31,24 +46,24 @@ function writeAll(all: Record<string, RememberedLoadSettings>) {
 }
 
 export function loadRememberedLoadSettings(
-  modelId: string,
+  key: string,
 ): RememberedLoadSettings | null {
-  return readAll()[modelId] ?? null;
+  return readAll()[key] ?? null;
 }
 
 export function saveRememberedLoadSettings(
-  modelId: string,
+  key: string,
   settings: RememberedLoadSettings,
 ) {
   const all = readAll();
-  all[modelId] = settings;
+  all[key] = settings;
   writeAll(all);
 }
 
-export function clearRememberedLoadSettings(modelId: string) {
+export function clearRememberedLoadSettings(key: string) {
   const all = readAll();
-  if (modelId in all) {
-    delete all[modelId];
+  if (key in all) {
+    delete all[key];
     writeAll(all);
   }
 }
diff --git a/studio/frontend/src/components/assistant-ui/reasoning.tsx b/studio/frontend/src/components/assistant-ui/reasoning.tsx
index a6c326c64e..c5b53577cb 100644
--- a/studio/frontend/src/components/assistant-ui/reasoning.tsx
+++ b/studio/frontend/src/components/assistant-ui/reasoning.tsx
@@ -20,7 +20,8 @@ import {
 } from "@assistant-ui/react";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { type VariantProps, cva } from "class-variance-authority";
-import { ChevronDownIcon, CopyIcon, LightbulbIcon } from "lucide-react";
+import { ChevronDownIcon, CopyIcon } from "lucide-react";
+import { BulbIcon } from "@/lib/bulb-icon";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { HugeiconsIcon } from "@hugeicons/react";
 import {
@@ -128,7 +129,7 @@ function ReasoningTrigger({
       )}
       {...props}
     >
-      <LightbulbIcon className="aui-reasoning-trigger-icon size-4 shrink-0" />
+      <BulbIcon className="aui-reasoning-trigger-icon size-4 shrink-0" />
       <span
         data-slot="reasoning-trigger-label"
         className="aui-reasoning-trigger-label-wrapper relative inline-block leading-none"
@@ -393,7 +394,8 @@ const ReasoningGroupImpl: ReasoningGroupComponent = ({
         <ReasoningTrigger
           className="min-w-0 flex-1"
           active={isReasoningStreaming}
-          duration={duration || persistedDuration}
+          // Prefer server timing when available.
+          duration={persistedDuration || duration}
         />
         <div className="flex w-16 shrink-0 justify-end">
           {isOpen && !isReasoningStreaming && (
diff --git a/studio/frontend/src/components/assistant-ui/thread.tsx b/studio/frontend/src/components/assistant-ui/thread.tsx
index 57e6d02f55..bad9a6b7f3 100644
--- a/studio/frontend/src/components/assistant-ui/thread.tsx
+++ b/studio/frontend/src/components/assistant-ui/thread.tsx
@@ -3043,6 +3043,7 @@ const ComposerToolsMenu: FC<{ side?: "top" | "bottom" }> = ({
           type="button"
           aria-label="Tools and attachments"
           className="unsloth-composer-plus"
+          data-tour="chat-plus-menu"
         >
           <PlusIcon className="size-[22px] stroke-[1.75px]" />
         </button>
diff --git a/studio/frontend/src/features/chat/api/chat-adapter.ts b/studio/frontend/src/features/chat/api/chat-adapter.ts
index a58cdd94d7..5e8fc9cb25 100644
--- a/studio/frontend/src/features/chat/api/chat-adapter.ts
+++ b/studio/frontend/src/features/chat/api/chat-adapter.ts
@@ -2627,6 +2627,17 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
                 continue;
               }
 
+              // Local GGUF sends server-timed reasoning duration. Guard the type
+              // so a malformed or proxied chunk (string/null/NaN duration) can
+              // never turn the label into NaN.
+              const reasoningMs = (
+                chunk as { _reasoningDurationMs?: number } | null | undefined
+              )?._reasoningDurationMs;
+              if (typeof reasoningMs === "number" && Number.isFinite(reasoningMs)) {
+                reasoningDuration = Math.max(0, Math.round(reasoningMs / 1000));
+                continue;
+              }
+
               // Diffusion frame: a transient canvas snapshot. Route it to the transient
               // store (the in-bubble renderer reads it) and skip it; it has no assistant
               // text, so it never enters the transcript or the counters below.
@@ -3175,6 +3186,7 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
               }
               const textParts = parseAssistantContent(cumulativeText);
 
+              // Fallback when no server-side reasoning_summary arrives.
               if (
                 textParts.some((part) => part.type === "reasoning") &&
                 !reasoningStartAt
@@ -3283,6 +3295,13 @@ export function createOpenAIStreamAdapter(): ChatModelAdapter {
           finalTokPerSec,
         );
 
+        // Finalize reasoning-only streams.
+        if (reasoningStartAt && !reasoningDuration) {
+          reasoningDuration = Math.max(
+            0,
+            Math.round((Date.now() - reasoningStartAt) / 1000),
+          );
+        }
         yield {
           content: [
             ...buildAssistantContent(cumulativeText),
diff --git a/studio/frontend/src/features/chat/api/chat-api.ts b/studio/frontend/src/features/chat/api/chat-api.ts
index 589fae5fe9..7112a7877c 100644
--- a/studio/frontend/src/features/chat/api/chat-api.ts
+++ b/studio/frontend/src/features/chat/api/chat-api.ts
@@ -902,6 +902,19 @@ export async function* streamChatCompletions(
         separatorIndex = buffer.search(/\r?\n\r?\n/);
         continue;
       }
+      // Relay server-side reasoning duration.
+      if (
+        parsed &&
+        typeof parsed === "object" &&
+        "type" in parsed &&
+        parsed.type === "reasoning_summary"
+      ) {
+        yield {
+          _reasoningDurationMs: (parsed as { duration_ms?: number }).duration_ms,
+        } as unknown as OpenAIChatChunk;
+        separatorIndex = buffer.search(/\r?\n\r?\n/);
+        continue;
+      }
       yield parsed as OpenAIChatChunk;
       separatorIndex = buffer.search(/\r?\n\r?\n/);
     }
diff --git a/studio/frontend/src/features/chat/chat-page.tsx b/studio/frontend/src/features/chat/chat-page.tsx
index daad2c4524..a288a60fb6 100644
--- a/studio/frontend/src/features/chat/chat-page.tsx
+++ b/studio/frontend/src/features/chat/chat-page.tsx
@@ -8,6 +8,10 @@ import {
   type ModelOption,
   ModelSelector,
 } from "@/components/assistant-ui/model-selector";
+import {
+  loadRememberedLoadSettings,
+  rememberedLoadSettingsKey,
+} from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import { ProjectComposer, Thread } from "@/components/assistant-ui/thread";
 import { CopyableErrorChip } from "@/components/ui/copyable-error-chip";
 import {
@@ -18,6 +22,10 @@ import {
 import { useSidebar } from "@/components/ui/sidebar";
 import { Tooltip, TooltipContent } from "@/components/ui/tooltip";
 import { useLatestRef } from "@/features/hub/hooks/use-latest-ref";
+import {
+  DOWNLOAD_KIND,
+  downloadManager,
+} from "@/features/hub/download-manager";
 import {
   type NativeIntent,
   NativeModelChip,
@@ -1093,6 +1101,11 @@ export function ChatPage({
   const abandonStaged = useCallback(() => {
     useChatRuntimeStore.getState().abandonStagedModel();
   }, []);
+  // Detach a staged pick on navigation without cancelling its download: the
+  // transfer keeps running in the manager and lands in cache, like Hub.
+  const detachStaged = useCallback(() => {
+    useChatRuntimeStore.getState().abandonStagedModel({ keepDownload: true });
+  }, []);
   // Tracks whether the chat page is still mounted, so a staged-load failure that
   // resolves after the user left chat doesn't resurrect the abandoned pick.
   const mountedRef = useRef(true);
@@ -1266,13 +1279,18 @@ export function ChatPage({
     selectModelRef.current = selectModel;
   }, [refresh, selectModel]);
   // Load a cached autoLoad pick once its download finishes. The sheet was never
-  // opened, so on a load failure just drop the orphaned staged knobs.
+  // opened, so on a load failure just drop the orphaned staged knobs. The knobs
+  // were already seeded on stage, so keepSpeculative only when a config was
+  // saved -- otherwise the standing speculative preference should win.
   autoLoadStagedRef.current = (pending) => {
+    const remembered = loadRememberedLoadSettings(
+      rememberedLoadSettingsKey(pending),
+    );
     void selectModel({
       ...pending,
       isDownloaded: true,
       forceReload: true,
-      keepSpeculative: false,
+      keepSpeculative: remembered != null,
       throwOnError: true,
     }).catch(() => {
       const store = useChatRuntimeStore.getState();
@@ -1620,8 +1638,8 @@ export function ChatPage({
     const prev = prevChatContextRef.current;
     prevChatContextRef.current = chatContextKey;
     if (prev === null || prev === chatContextKey) return;
-    abandonStaged();
-  }, [chatContextKey, abandonStaged]);
+    detachStaged();
+  }, [chatContextKey, detachStaged]);
 
   const hasActiveModel = Boolean(inferenceParams.checkpoint);
   // Load immediately, or — when "Load on selection" is off — stage the pick so
@@ -1639,25 +1657,81 @@ export function ChatPage({
         (!hasGgufSource(selection) && !wantManagerDownload) ||
         (store.loadOnSelection && selection.isDownloaded)
       ) {
-        // Abandon any staged pick first so its edited knobs (e.g. a custom
+        // Detach any staged pick first so its edited knobs (e.g. a custom
         // context length) don't leak into this immediate load -- resolveLoad
-        // reads customContextLength before checking the target is GGUF.
-        abandonStaged();
-        await selectModel(selection);
+        // reads customContextLength before checking the target is GGUF. Detach
+        // (not abandon) keeps its download running.
+        detachStaged();
+        // Load-on-selection skips the sheet, so seed the saved knobs here the
+        // way the sheet's restore effect would; the switch would otherwise reset
+        // the remembered speculative choice (keepSpeculative below prevents it).
+        const remembered = hasGgufSource(selection)
+          ? loadRememberedLoadSettings(rememberedLoadSettingsKey(selection))
+          : null;
+        if (remembered) store.applyRememberedLoadSettings(remembered);
+        await selectModel(
+          remembered ? { ...selection, keepSpeculative: true } : selection,
+        );
         return;
       }
-      // Refuse staging while a load is in flight (it would be silently dropped);
-      // the immediate-load branch above is already guarded in selectModel.
+      // Loads can't queue behind each other, but a download is independent: if
+      // the pick needs downloading, start it in the manager so it runs alongside
+      // the load. Nothing to download (already on device) just waits.
       if (store.modelLoading) {
-        toast.info("Another model is already loading", {
-          description: "Wait for it to finish or cancel it first.",
-        });
+        // Both an uncached non-GGUF snapshot (wantManagerDownload) and an
+        // uncached remote GGUF quant download through the manager, so either can
+        // run in the background while another model loads. wantManagerDownload
+        // excludes GGUF by design, so the GGUF case is checked separately.
+        const wantBackgroundDownload =
+          wantManagerDownload ||
+          (selection.source === "hub" &&
+            hasGgufSource(selection) &&
+            !selection.isDownloaded);
+        // The model currently loading already downloads as part of its own load
+        // (the /load flow fetches before setting the checkpoint), so re-picking
+        // it must not kick off a second transfer against the same cache.
+        const isLoadingThisPick =
+          !!loadingModel &&
+          normalizeModelRef(loadingModel.id) ===
+            normalizeModelRef(selection.id) &&
+          (loadingModel.ggufVariant ?? null) === (selection.ggufVariant ?? null);
+        if (isLoadingThisPick) {
+          toast.info("This model is already loading", {
+            description: "It's downloading as part of the load in progress.",
+          });
+        } else if (wantBackgroundDownload) {
+          // Only claim the download started once a job is actually created. A
+          // transport conflict records state that is only resolvable from the
+          // Hub download card, so point the user there instead of showing a
+          // success toast for a transfer that never began; "busy" and "error"
+          // already surface their own toasts.
+          const outcome = await downloadManager.requestStart({
+            kind: DOWNLOAD_KIND.MODEL,
+            repoId: selection.id,
+            variant: selection.ggufVariant ?? null,
+            expectedBytes: selection.expectedBytes ?? 0,
+          });
+          if (outcome === "started") {
+            toast.info("Downloading in the background", {
+              description:
+                "It'll be ready to load once the current model finishes.",
+            });
+          } else if (outcome === "conflict") {
+            toast.info("Resume this download from the Hub", {
+              description:
+                "An earlier partial download used a different transport. Open the Hub tab to resume or restart it.",
+            });
+          }
+        } else {
+          toast.info("Another model is already loading", {
+            description: "Wait for it to finish or cancel it first.",
+          });
+        }
         return;
       }
-      // Tear down any existing staged pick first so its in-flight download is
-      // cancelled, not left running after we rebind to the new pick. With the
-      // toggle on, autoLoad downloads silently then loads; off stages for the sheet.
-      abandonStaged();
+      // Detach the prior staged pick (keeping its download) before rebinding, so
+      // a second pick downloads alongside the first instead of cancelling it.
+      detachStaged();
       store.stageModel({
         id: selection.id,
         isLora: selection.isLora,
@@ -1670,7 +1744,7 @@ export function ChatPage({
         autoLoad: store.loadOnSelection,
       });
     },
-    [abandonStaged, selectModel],
+    [detachStaged, selectModel, loadingModel],
   );
   const loadNativeModelIntent = useCallback(
     async (intent: NativeIntent, loadingDescription: string) => {
@@ -2452,7 +2526,6 @@ export function ChatPage({
                     onClick={() => setSettingsOpen(true)}
                     className="flex h-[34px] w-[34px] translate-x-[2px] cursor-pointer items-center justify-center rounded-full text-nav-fg transition-colors hover:bg-nav-surface-hover hover:text-black dark:hover:text-white focus-visible:outline-none focus-visible:ring-2 focus-visible:ring-ring"
                     aria-label="Open run settings"
-                    data-tour="chat-settings"
                   >
                     <HugeiconsIcon
                       icon={LayoutAlignRightIcon}
diff --git a/studio/frontend/src/features/chat/chat-settings-sheet.tsx b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
index bd315e19d4..bc73e1ddf2 100644
--- a/studio/frontend/src/features/chat/chat-settings-sheet.tsx
+++ b/studio/frontend/src/features/chat/chat-settings-sheet.tsx
@@ -21,6 +21,7 @@ import { Checkbox } from "@/components/ui/checkbox";
 import {
   clearRememberedLoadSettings,
   loadRememberedLoadSettings,
+  rememberedLoadSettingsKey,
   saveRememberedLoadSettings,
 } from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import {
@@ -579,6 +580,9 @@ export function ChatSettingsPanel({
   );
   const kvCacheDtype = useChatRuntimeStore((s) => s.kvCacheDtype);
   const setKvCacheDtype = useChatRuntimeStore((s) => s.setKvCacheDtype);
+  const applyRememberedLoadSettings = useChatRuntimeStore(
+    (s) => s.applyRememberedLoadSettings,
+  );
   const loadedKvCacheDtype = useChatRuntimeStore((s) => s.loadedKvCacheDtype);
   const tensorParallel = useChatRuntimeStore((s) => s.tensorParallel);
   const setTensorParallel = useChatRuntimeStore((s) => s.setTensorParallel);
@@ -613,25 +617,16 @@ export function ChatSettingsPanel({
   // the saved per-model settings on stage, so the sheet opens with what was used
   // last time; the tick reflects whether a saved entry exists.
   const [remember, setRemember] = useState(false);
-  const pendingId = pendingSelection?.id ?? null;
+  // Keyed per quant: a different variant of the same repo has its own settings.
+  const pendingKey = pendingSelection
+    ? rememberedLoadSettingsKey(pendingSelection)
+    : null;
   useEffect(() => {
-    if (!pendingId) return;
-    const saved = loadRememberedLoadSettings(pendingId);
+    if (!pendingKey) return;
+    const saved = loadRememberedLoadSettings(pendingKey);
     setRemember(saved != null);
-    if (!saved) return;
-    setCustomContextLength(saved.contextLength);
-    setKvCacheDtype(saved.kvCacheDtype);
-    setSpeculativeType(saved.speculativeType ?? "auto");
-    setSpecDraftNMax(saved.specDraftNMax);
-    setTensorParallel(saved.tensorParallel);
-  }, [
-    pendingId,
-    setCustomContextLength,
-    setKvCacheDtype,
-    setSpeculativeType,
-    setSpecDraftNMax,
-    setTensorParallel,
-  ]);
+    if (saved) applyRememberedLoadSettings(saved);
+  }, [pendingKey, applyRememberedLoadSettings]);
   // While staging, the sheet reflects the STAGED model, so its header context
   // takes precedence over the loaded model's (which may differ or be larger).
   const baseContext = pendingIsGguf ? stagedContextLength : ggufContextLength;
@@ -1213,9 +1208,11 @@ export function ChatSettingsPanel({
                       type="button"
                       onClick={() => {
                         // Persist (or clear) this model's load knobs before loading.
-                        // Save the explicit context override only (null = auto), so
-                        // restoring never forces the native context into an OOM.
-                        const pid = pendingSelection?.id;
+                        // Context is stored as the override (null = auto), never the
+                        // resolved native value, so restoring can't force an OOM.
+                        const pid = pendingSelection
+                          ? rememberedLoadSettingsKey(pendingSelection)
+                          : null;
                         if (pid) {
                           if (remember) {
                             saveRememberedLoadSettings(pid, {
@@ -1775,7 +1772,9 @@ export function ChatSettingsPanel({
             <SheetTitle>Run settings</SheetTitle>
             <SheetDescription>Chat inference settings</SheetDescription>
           </SheetHeader>
-          <div className="flex h-full flex-col">{settingsContent}</div>
+          <div data-tour="chat-settings" className="flex h-full flex-col">
+            {settingsContent}
+          </div>
         </SheetContent>
       </Sheet>
     );
@@ -1783,6 +1782,7 @@ export function ChatSettingsPanel({
 
   return (
     <aside
+      data-tour="chat-settings"
       className={`relative z-50 shrink-0 h-full overflow-hidden bg-panel-surface text-panel-surface-fg font-heading ${open ? "w-[17rem] border-l border-sidebar-border" : "w-0"}`}
     >
       <div className="h-full w-full">{settingsContent}</div>
diff --git a/studio/frontend/src/features/chat/shared-composer.tsx b/studio/frontend/src/features/chat/shared-composer.tsx
index f35ccc6149..912d92003b 100644
--- a/studio/frontend/src/features/chat/shared-composer.tsx
+++ b/studio/frontend/src/features/chat/shared-composer.tsx
@@ -7,6 +7,7 @@ import {
   thinkToggleAriaLabel,
 } from "@/components/assistant-ui/think-aria-label";
 import { Button } from "@/components/ui/button";
+import { BulbIcon } from "@/lib/bulb-icon";
 import { Tick02Icon } from "@/lib/tick-icon";
 import { cn } from "@/lib/utils";
 import {
@@ -155,20 +156,6 @@ const MicIcon: FC<{ className?: string }> = ({ className }) => (
   </svg>
 );
 
-const BulbIcon: FC<{ className?: string }> = ({ className }) => (
-  <svg
-    className={className}
-    viewBox="-10.24 -10.24 1044.48 1044.48"
-    fill="currentColor"
-    stroke="currentColor"
-    strokeWidth={16.384}
-    xmlns="http://www.w3.org/2000/svg"
-    aria-hidden={true}
-  >
-    <path d="M511.984 0c-198.032 0-353.12 161.104-353.12 359.136 0 149.2 73.28 220.256 131.185 272.128 37.28 33.424 62.368 53.552 62.368 78.352v54.255c0 1.392.193 2.752.368 4.128h-.72v92.624c.016 97.712 63.2 163.376 161.072 163.376 94.464 0 158.944-65.664 158.944-163.376V768h-.928c.176-1.376.416-2.736.416-4.128v-54.255c0-37.76 28.032-60.592 70.528-97.696 57.504-50.208 123.023-112.688 123.023-252.784C865.136 161.104 710.016 0 511.983 0zm-1.215 960c-59.904 0-94.689-37.152-94.689-99.376l-.463-42.672C438.64 825.824 470 832 512 832c41.424 0 72.848-6.624 96.08-14.768v43.392c0 63.152-35.247 99.376-97.312 99.376zm189.248-396.288c-43.472 37.968-92.433 77.216-92.433 145.904v40.432c-15.183 8.48-43.183 18.56-96.127 18.56-55.569 0-81.92-9.856-95.024-17.473V709.6c0-54.608-42.688-89.297-83.68-126.017-54.32-48.672-109.873-103.84-109.873-224.464-.015-162.72 126.385-295.12 289.104-295.12 162.752 0 289.152 132.4 289.152 295.137 0 111.024-48.463 158.576-101.12 204.576z" />
-  </svg>
-);
-
 function isNativeComposing(event: Event) {
   return "isComposing" in event && (event as InputEvent).isComposing === true;
 }
diff --git a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
index 4cca3dc2e3..5b2dae5e08 100644
--- a/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
+++ b/studio/frontend/src/features/chat/stores/chat-runtime-store.ts
@@ -1,6 +1,7 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import type { RememberedLoadSettings } from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import { cancelStagedModelDownload } from "@/features/hub";
 import { toast } from "@/lib/toast";
 import { create } from "zustand";
@@ -770,6 +771,10 @@ type ChatRuntimeStore = {
    *  start each deferred-staging session clean so one staged pick's settings
    *  don't leak onto the next. */
   resetModelSettingsToLoaded: () => void;
+  /** Seed the editable load knobs from a model's remembered settings. Shared by
+   *  the settings sheet's restore effect and the "Load on selection" paths,
+   *  which skip the sheet but must still honor a saved config. */
+  applyRememberedLoadSettings: (settings: RememberedLoadSettings) => void;
   setTensorParallel: (value: boolean) => void;
   setLoadOnSelection: (value: boolean) => void;
   setExpandQuantizations: (value: boolean) => void;
@@ -778,9 +783,10 @@ type ChatRuntimeStore = {
   /** Stage a pick for a deferred load: revert knobs to the loaded baseline,
    *  record the selection, and open the settings sheet. */
   stageModel: (selection: PendingModelSelection) => void;
-  /** Abandon a staged pick without loading: revert the knobs to the loaded
-   *  baseline and clear the pending selection. */
-  abandonStagedModel: () => void;
+  /** Abandon a staged pick without loading: revert knobs to the loaded baseline
+   *  and clear the pending selection. Cancels its in-flight download too, unless
+   *  `keepDownload` is set (navigation keeps the transfer running, like Hub). */
+  abandonStagedModel: (opts?: { keepDownload?: boolean }) => void;
   setCustomContextLength: (v: number | null) => void;
   setChatTemplateOverride: (template: string | null) => void;
   setPendingAudio: (base64: string, name: string) => void;
@@ -1527,6 +1533,16 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
   setSpecDraftNMax: (specDraftNMax) => set({ specDraftNMax }),
   setTensorParallel: (tensorParallel) => set({ tensorParallel }),
   resetModelSettingsToLoaded: () => set((s) => loadedBaselineSettings(s)),
+  applyRememberedLoadSettings: (settings) =>
+    // Coalesce every field: a blob persisted by an older/newer build can omit
+    // keys, and a raw spread would push `undefined` into fields typed non-null.
+    set({
+      customContextLength: settings.contextLength ?? null,
+      kvCacheDtype: settings.kvCacheDtype ?? null,
+      speculativeType: settings.speculativeType ?? "auto",
+      specDraftNMax: settings.specDraftNMax ?? null,
+      tensorParallel: settings.tensorParallel ?? false,
+    }),
   setLoadOnSelection: (loadOnSelection) => {
     saveBool(CHAT_LOAD_ON_SELECTION_KEY, loadOnSelection);
     set({ loadOnSelection });
@@ -1544,15 +1560,9 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
     // Refuse staging mid-load: post-load cleanup would silently drop the queued
     // pick. stageOrLoad toasts first for callers that can.
     if (get().modelLoading) return;
+    // Rebinding to a new pick keeps the prior pick's download running so the
+    // user can queue multiple downloads at once (Hub-style).
     set((s) => {
-      if (
-        s.pendingSelection &&
-        (s.pendingSelection.id !== selection.id ||
-          (s.pendingSelection.ggufVariant ?? null) !==
-            (selection.ggufVariant ?? null))
-      ) {
-        cancelStagedModelDownload(s.pendingSelection);
-      }
       return {
         ...loadedBaselineSettings(s),
         pendingSelection: selection,
@@ -1566,14 +1576,13 @@ export const useChatRuntimeStore = create<ChatRuntimeStore>((set, get) => ({
       };
     });
   },
-  abandonStagedModel: () => {
+  abandonStagedModel: (opts) => {
     const { pendingSelection } = get();
     if (!pendingSelection) return;
-    // Cancel the staged pick's in-flight download so it doesn't keep running
-    // after the staging UI is gone. Centralized here so every abandon path
-    // (sheet close, thread switch, route exit, new chat) cancels it, including
-    // root-level callers that have no access to the useRepoDownload hook.
-    cancelStagedModelDownload(pendingSelection);
+    // Cancel the staged pick's in-flight download (centralized for every abandon
+    // path: sheet close, thread switch, route exit, new chat). `keepDownload`
+    // opts out so navigation leaves the transfer running, like a Hub download.
+    if (!opts?.keepDownload) cancelStagedModelDownload(pendingSelection);
     set((s) => ({ ...loadedBaselineSettings(s), pendingSelection: null }));
   },
   setCustomContextLength: (customContextLength) => set({ customContextLength }),
diff --git a/studio/frontend/src/features/chat/tour/steps.tsx b/studio/frontend/src/features/chat/tour/steps.tsx
index cbbb7bae4f..84e3fe5950 100644
--- a/studio/frontend/src/features/chat/tour/steps.tsx
+++ b/studio/frontend/src/features/chat/tour/steps.tsx
@@ -27,20 +27,21 @@ export function buildChatTourSteps({
       title: "Pick a model",
       body: (
         <>
-          This selects what’s loaded for inference. Hub = base models. Fine-tuned
-          = trained Unsloth outputs, including LoRA adapters and full finetunes.
+          Selects what’s loaded for inference. Recommended is Unsloth’s curated
+          base models; On Device is your downloads and fine-tuned outputs (LoRA
+          adapters and full finetunes).
         </>
       ),
     },
     {
       id: "model-tabs",
       target: "chat-model-selector-popover",
-      title: "Two tabs",
+      title: "Find a model",
       body: (
         <>
-          Hub: search Hugging Face models. Fine-tuned: local Unsloth outputs you’ve
-          trained or exported. If results look off, compare base vs fine-tuned
-          outputs to see what changed.
+          Search Unsloth’s models, or hit Search Hub for all of Hugging Face.
+          Switch Recommended and On Device, filter by format, and sort by
+          trending or recent. An OOM tag means it won’t fit in your VRAM.
         </>
       ),
       onEnter: openModelSelector,
@@ -59,6 +60,18 @@ export function buildChatTourSteps({
       onEnter: openSettings,
       onExit: closeSettings,
     },
+    {
+      id: "plus-menu",
+      target: "chat-plus-menu",
+      title: "The + menu",
+      body: (
+        <>
+          Everything else lives here: attach photos and files, reuse saved
+          prompts, toggle tools and MCP, start a side-by-side compare, and
+          export the chat.
+        </>
+      ),
+    },
   ];
 
   if (canCompare) {
diff --git a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
index c8adfb54f5..48f7fcffaa 100644
--- a/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
+++ b/studio/frontend/src/features/hub/catalog/models-toolbar.tsx
@@ -252,7 +252,7 @@ export const ModelsToolbar = memo(function ModelsToolbar({
                 ? `Search on-device ${isDataset ? "datasets" : "models"}`
                 : isDataset
                   ? "Search datasets"
-                  : "Search models"
+                  : "Search all models"
             }
             className={cn(
               "field-soft h-9 rounded-full !border-0 pl-10 text-[13px] placeholder:text-muted-foreground/80 focus-visible:!ring-0",
diff --git a/studio/frontend/src/features/hub/catalog/owner-avatar.tsx b/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
index 82a1f150d2..852c09ab5f 100644
--- a/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
+++ b/studio/frontend/src/features/hub/catalog/owner-avatar.tsx
@@ -19,24 +19,6 @@ const SIZES: Record<AvatarSize, string> = {
   lg: "size-12 rounded-[15px] text-[16px]",
 };
 
-// Unsloth's own uploads (no upstream provider match) show the bundled Unsloth
-// brand avatar instead of a colored-initial tile, so they read as Unsloth even
-// in virtualized rows that never fetch the HF profile picture.
-const UNSLOTH_OWNER_LOGO: ProviderLogo = {
-  id: "unsloth",
-  name: "Unsloth",
-  logoPath: "/circle-logo-small.png",
-  treatment: "original",
-  background: "transparent",
-  fit: "cover",
-  // Used directly for the unsloth owner, never via prefix matching.
-  prefixes: [],
-};
-
-function isUnslothOwner(owner: string): boolean {
-  return owner.trim().toLowerCase() === "unsloth";
-}
-
 const AVATAR_IMAGE_RETRY_BASE_MS = 60_000;
 const AVATAR_IMAGE_RETRY_MAX_MS = 30 * 60_000;
 
@@ -86,15 +68,6 @@ export function OwnerAvatar({
       />
     );
   }
-  if (isUnslothOwner(owner)) {
-    return (
-      <ProviderLogoTile
-        provider={UNSLOTH_OWNER_LOGO}
-        size={size}
-        className={className}
-      />
-    );
-  }
   return (
     <DefaultAvatar
       owner={owner}
@@ -114,9 +87,6 @@ export function useAvatarImageUrl(
   if (provider) {
     return provider.treatment === "original" ? provider.logoPath : null;
   }
-  if (isUnslothOwner(owner)) {
-    return UNSLOTH_OWNER_LOGO.logoPath;
-  }
   return remoteUrl;
 }
 
diff --git a/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx b/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
index fed8568f99..5f36f5d031 100644
--- a/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
+++ b/studio/frontend/src/features/hub/catalog/owner-scope-toggle.tsx
@@ -28,8 +28,8 @@ export function OwnerScopeToggle({
       onValueChange={onChange}
       ariaLabel="Publisher scope"
       align="end"
-      // Extra gap so the chevron sits a touch further from the label.
-      className="h-8 gap-1.5 text-[11.5px]"
+      // Extra gap before the chevron; min-width keeps the pill readable.
+      className="h-8 min-w-[96px] gap-1.5 text-[11.5px]"
     />
   );
 }
diff --git a/studio/frontend/src/features/hub/download-manager/transport-conflict.ts b/studio/frontend/src/features/hub/download-manager/transport-conflict.ts
index 9125abed70..3267a6ae05 100644
--- a/studio/frontend/src/features/hub/download-manager/transport-conflict.ts
+++ b/studio/frontend/src/features/hub/download-manager/transport-conflict.ts
@@ -80,24 +80,50 @@ async function activeSiblingTransport(
   return null;
 }
 
+// Outcome of a start request so callers can tell whether a transfer for this
+// exact request is actually live before telling the user it began. "started"
+// means a running/cancelling job exists for this key (a fresh start or an
+// already-active one). "conflict" means a transport partial conflict was
+// recorded and must be resolved from the Hub download card; "busy" means the
+// repo is occupied by a sibling variant/snapshot/pending start that is not this
+// transfer; "error" means the start failed or was refused.
+export type DownloadStartOutcome = "started" | "conflict" | "busy" | "error";
+
+// A start can no-op without throwing: the backend can refuse it (startJob
+// finalizes "error"), startJob's peer guard can skip it, or
+// hasActiveOrPendingStart can trip on a snapshot/peer/pending that is not this
+// request. Derive the outcome from the actual job state of this exact key so
+// callers never claim a download began when it did not.
+function isJobActiveFor(req: DownloadRequest): boolean {
+  const job = getState().jobs[jobKeyOf(req.kind, req.repoId, req.variant)];
+  return Boolean(job && ACTIVE_STATES.has(job.state));
+}
+
 async function runWithPendingStartGuard(
   req: DownloadRequest,
-  action: () => Promise<void>,
-): Promise<void> {
+  action: () => Promise<DownloadStartOutcome>,
+): Promise<DownloadStartOutcome> {
   const startKey = pendingStartKey(req);
-  if (hasActiveOrPendingStart(req)) return;
+  // Already active or pending for the repo: only report "started" when this
+  // exact request is the live transfer; a peer/snapshot/pending start has not.
+  if (hasActiveOrPendingStart(req)) {
+    return isJobActiveFor(req) ? "started" : "busy";
+  }
   runtimeRegistry.pendingStartRepoKeys.add(startKey);
   try {
-    await action();
+    return await action();
   } catch (error) {
     reportConflictStartError(error);
+    return "error";
   } finally {
     runtimeRegistry.pendingStartRepoKeys.delete(startKey);
   }
 }
 
-export async function requestStart(req: DownloadRequest): Promise<void> {
-  await runWithPendingStartGuard(req, async () => {
+export async function requestStart(
+  req: DownloadRequest,
+): Promise<DownloadStartOutcome> {
+  return runWithPendingStartGuard(req, async () => {
     let mode: TransportMode = getTransportMode();
     try {
       mode = await effectiveTransportMode(mode);
@@ -119,7 +145,7 @@ export async function requestStart(req: DownloadRequest): Promise<void> {
               ? "This repository is currently downloading with Xet. Switch to Xet or wait for it to finish."
               : "This repository is currently downloading with HTTP. Switch to HTTP or wait for it to finish.",
         });
-        return;
+        return "busy";
       }
     } catch (err) {
       console.warn("Active download transport check failed.", err);
@@ -139,7 +165,7 @@ export async function requestStart(req: DownloadRequest): Promise<void> {
           },
           pending: req,
         });
-        return;
+        return "conflict";
       }
       if (status.has_partial && !status.last_transport) {
         toast.info("Restarting this download", {
@@ -163,7 +189,7 @@ export async function requestStart(req: DownloadRequest): Promise<void> {
             "Starting with HTTP so an existing partial is not discarded. Switch transport to retry with Xet.",
         });
         await startJob(req, { useXet: false });
-        return;
+        return isJobActiveFor(req) ? "started" : "error";
       }
       toast.warning("Couldn't verify existing partial download", {
         description:
@@ -171,6 +197,7 @@ export async function requestStart(req: DownloadRequest): Promise<void> {
       });
     }
     await startJob(req, { useXet: mode === TRANSPORT.XET });
+    return isJobActiveFor(req) ? "started" : "error";
   });
 }
 
@@ -178,22 +205,24 @@ export function resumeConflict(conflictKey: string): void {
   const entry = getState().conflicts[conflictKey];
   if (!entry) return;
   setConflict(conflictKey, null);
-  void runWithPendingStartGuard(entry.pending, () =>
-    startJob(entry.pending, {
+  void runWithPendingStartGuard(entry.pending, async () => {
+    await startJob(entry.pending, {
       useXet: entry.info.previous === TRANSPORT.XET,
-    }),
-  );
+    });
+    return "started";
+  });
 }
 
 export function restartConflict(conflictKey: string): void {
   const entry = getState().conflicts[conflictKey];
   if (!entry) return;
   setConflict(conflictKey, null);
-  void runWithPendingStartGuard(entry.pending, () =>
-    startJob(entry.pending, {
+  void runWithPendingStartGuard(entry.pending, async () => {
+    await startJob(entry.pending, {
       useXet: entry.info.next === TRANSPORT.XET,
-    }),
-  );
+    });
+    return "started";
+  });
 }
 
 export function cancelConflict(conflictKey: string): void {
diff --git a/studio/frontend/src/features/hub/download-manager/use-repo-download.ts b/studio/frontend/src/features/hub/download-manager/use-repo-download.ts
index 46b5fae6bb..9b9f33d8a2 100644
--- a/studio/frontend/src/features/hub/download-manager/use-repo-download.ts
+++ b/studio/frontend/src/features/hub/download-manager/use-repo-download.ts
@@ -120,8 +120,10 @@ export function useRepoDownload(config: RepoDownloadConfig): DownloadJob {
   );
 
   const requestStartDownload = useCallback(
-    (variant: string | null, expectedBytes: number) => {
-      return downloadManager.requestStart({
+    async (variant: string | null, expectedBytes: number) => {
+      // This surface renders the conflict resolver (transportConflict), so the
+      // start outcome is handled by the card UI; the awaited result is ignored.
+      await downloadManager.requestStart({
         kind,
         repoId,
         variant,
diff --git a/studio/frontend/src/features/hub/hub-page.tsx b/studio/frontend/src/features/hub/hub-page.tsx
index 943aca5e9a..c696862283 100644
--- a/studio/frontend/src/features/hub/hub-page.tsx
+++ b/studio/frontend/src/features/hub/hub-page.tsx
@@ -1,6 +1,10 @@
 // SPDX-License-Identifier: AGPL-3.0-only
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
+import {
+  loadRememberedLoadSettings,
+  rememberedLoadSettingsKey,
+} from "@/components/assistant-ui/model-selector/remembered-load-settings";
 import { useHubInventory } from "@/features/hub/inventory";
 import { useDebouncedValue } from "@/hooks/use-debounced-value";
 import { useGpuInfo } from "@/hooks/use-gpu-info";
@@ -14,7 +18,10 @@ import { useHubInfiniteScroll } from "@/features/hub/hooks/use-hub-infinite-scro
 import { ggufVariantsMatch, modelIdsMatch } from "@/features/hub/lib/model-identity";
 import { cn } from "@/lib/utils";
 import { usePlatformStore } from "@/config/env";
-import { useHfTokenStore } from "@/features/hub/stores/hf-token-store";
+import {
+  hfApiToken,
+  useHfTokenStore,
+} from "@/features/hub/stores/hf-token-store";
 import {
   getInferenceStatus,
   isExternalModelId,
@@ -90,18 +97,19 @@ const ALL_MODELS_VIEW_STORAGE_KEY = "unsloth.hub.allModelsView";
 const INVENTORY_SORT_STORAGE_KEY = "unsloth.hub.inventorySort";
 const OWNER_SCOPE_STORAGE_KEY = "unsloth.hub.ownerScope";
 
-/** Discover browsing scope: only the unsloth org (default) or the whole Hub. */
+/** Discover browsing scope: the whole Hub (default) or only the unsloth org. */
 export type OwnerScope = "unsloth" | "all";
 
 function readOwnerScopePreference(): OwnerScope {
   if (typeof window === "undefined") {
-    return "unsloth";
+    return "all";
   }
   try {
     const value = window.localStorage.getItem(OWNER_SCOPE_STORAGE_KEY);
-    return value === "all" ? "all" : "unsloth";
+    // Default to the whole Hub; only honor an explicit "unsloth" preference.
+    return value === "unsloth" ? "unsloth" : "all";
   } catch {
-    return "unsloth";
+    return "all";
   }
 }
 
@@ -550,6 +558,7 @@ export function ModelsPage() {
   const deferredDebouncedQuery = useDeferredValue(debouncedQuery);
   const hfToken = useHfTokenStore((s) => s.token);
   const debouncedHfToken = useDebouncedValue(hfToken, 500);
+  const apiHfToken = hfApiToken(debouncedHfToken);
   const deferredFormatFilter = useDeferredValue(formatFilter);
   const deferredCapabilityFilter = useDeferredValue(capabilityFilter);
 
@@ -604,7 +613,7 @@ export function ModelsPage() {
     handleRetrySearch,
   } = useDiscoverSearch({
     debouncedQuery,
-    accessToken: debouncedHfToken || undefined,
+    accessToken: apiHfToken,
     isDiscoverTab,
     isDatasetMode,
     sortBy: effectiveSort,
@@ -618,7 +627,7 @@ export function ModelsPage() {
     channelId: isChannelListMode ? activeChannelId : null,
     results,
     isLoading,
-    accessToken: debouncedHfToken || undefined,
+    accessToken: apiHfToken,
   });
 
   const {
@@ -701,7 +710,7 @@ export function ModelsPage() {
   const listRows = filteredDiscoverRows;
 
   const hubFeed = useHubFeed({
-    accessToken: debouncedHfToken || undefined,
+    accessToken: apiHfToken,
     online,
     enabled: isFeedMode,
     deviceType,
@@ -740,6 +749,23 @@ export function ModelsPage() {
     () => (isDiscoverTab ? [] : tokenizeQuery(deferredDebouncedQuery)),
     [isDiscoverTab, deferredDebouncedQuery],
   );
+  // Hide infra models (e.g. the RAG embedder bge-small-en-v1.5) from the On
+  // Device list like Discover, but reveal a row when a query matches it so the
+  // user can confirm it is already downloaded.
+  const isVisibleInventoryRow = useCallback(
+    (row: CachedInventoryRow | LocalInventoryRow) =>
+      // Local rows can have a null repoId and an id that is a hash rather than
+      // the file path/name, so also check path/title (the backend's
+      // _is_hidden_model checks the on-disk path for the same reason).
+      !isHiddenModelId(
+        row.id,
+        row.repoId,
+        row.kind !== "cache" ? row.path : undefined,
+        row.kind !== "cache" ? row.title : undefined,
+      ) ||
+      (inventoryTokens.length > 0 && inventoryRowMatches(row, inventoryTokens)),
+    [inventoryTokens],
+  );
   // Format filter is a deliberate scope narrowing, so hard-filter it out. The
   // text query instead drives dim-not-filter on On Device (see ModelsCatalog) so
   // selection survives typing; matching rows are partitioned to the top.
@@ -748,12 +774,22 @@ export function ModelsPage() {
       partitionByMatch(
         effectiveCachedRows.filter(
           (row) =>
+            // Hidden-model filtering is model-only; datasets bypass it (and the
+            // format filter) the way Discover does, so a dataset whose
+            // id/title/path happens to contain an infra needle is not dropped.
             isDatasetMode ||
-            matchesFormat(row.modelFormat, deferredFormatFilter),
+            (matchesFormat(row.modelFormat, deferredFormatFilter) &&
+              isVisibleInventoryRow(row)),
         ),
         inventoryTokens,
       ),
-    [effectiveCachedRows, isDatasetMode, deferredFormatFilter, inventoryTokens],
+    [
+      effectiveCachedRows,
+      isDatasetMode,
+      deferredFormatFilter,
+      inventoryTokens,
+      isVisibleInventoryRow,
+    ],
   );
 
   const filteredLocalRows = useMemo(
@@ -761,12 +797,42 @@ export function ModelsPage() {
       partitionByMatch(
         effectiveLocalRows.filter(
           (row) =>
+            // Hidden-model filtering is model-only; datasets bypass it (and the
+            // format filter) the way Discover does, so a dataset whose
+            // id/title/path happens to contain an infra needle is not dropped.
             isDatasetMode ||
-            matchesFormat(row.modelFormat, deferredFormatFilter),
+            (matchesFormat(row.modelFormat, deferredFormatFilter) &&
+              isVisibleInventoryRow(row)),
         ),
         inventoryTokens,
       ),
-    [effectiveLocalRows, isDatasetMode, deferredFormatFilter, inventoryTokens],
+    [
+      effectiveLocalRows,
+      isDatasetMode,
+      deferredFormatFilter,
+      inventoryTokens,
+      isVisibleInventoryRow,
+    ],
+  );
+
+  // Header tallies exclude infra/hidden models so the count matches the On
+  // Device list (a fresh install with only the bge embedder cached reads 0,
+  // not 1 over an empty list). Reuse isVisibleInventoryRow so a hidden row
+  // revealed by an active search is counted too, and datasets (never infra)
+  // keep their full count, mirroring the row filter above.
+  const visibleCachedCount = useMemo(
+    () =>
+      effectiveCachedRows.filter(
+        (row) => isDatasetMode || isVisibleInventoryRow(row),
+      ).length,
+    [effectiveCachedRows, isDatasetMode, isVisibleInventoryRow],
+  );
+  const visibleLocalCount = useMemo(
+    () =>
+      effectiveLocalRows.filter(
+        (row) => isDatasetMode || isVisibleInventoryRow(row),
+      ).length,
+    [effectiveLocalRows, isDatasetMode, isVisibleInventoryRow],
   );
 
   const filterResetSignature = useMemo(
@@ -850,7 +916,7 @@ export function ModelsPage() {
     filteredCachedRows,
     filteredLocalRows,
     results: selectionResults,
-    accessToken: debouncedHfToken || undefined,
+    accessToken: apiHfToken,
     online,
   });
 
@@ -1026,11 +1092,32 @@ export function ModelsPage() {
         openNewChat();
         return;
       }
+      // Detach any leftover staged pick first so its edited knobs (e.g. a custom
+      // context length) don't leak into this load -- mirrors the chat page's
+      // detachStaged(); keepDownload keeps any staged download running.
+      useChatRuntimeStore.getState().abandonStagedModel({ keepDownload: true });
+      // Load-on-selection skips the chat sheet, so seed this GGUF pick's saved
+      // load knobs here the way the sheet's restore effect would; otherwise the
+      // remembered config is silently ignored on the Hub run path. keepSpeculative
+      // then honors the restored speculative choice across the switch.
+      const remembered =
+        opts.ggufVariant != null || selectedModel.isGguf
+          ? loadRememberedLoadSettings(
+              rememberedLoadSettingsKey({
+                id: runId,
+                ggufVariant: opts.ggufVariant,
+              }),
+            )
+          : null;
+      if (remembered) {
+        useChatRuntimeStore.getState().applyRememberedLoadSettings(remembered);
+      }
       void selectModel({
         id: runId,
         ggufVariant: opts.ggufVariant,
         isDownloaded,
         expectedBytes: opts.expectedBytes,
+        keepSpeculative: remembered != null,
         throwOnError: true,
       })
         .then(() => {
@@ -1315,15 +1402,15 @@ export function ModelsPage() {
     return (
       <HubListHeader
         title="On device"
-        count={effectiveCachedRows.length + effectiveLocalRows.length}
+        count={visibleCachedCount + visibleLocalCount}
         view={allModelsView}
         onViewChange={setAllModelsView}
         actions={sortControl}
       />
     );
   }, [
-    effectiveCachedRows.length,
-    effectiveLocalRows.length,
+    visibleCachedCount,
+    visibleLocalCount,
     allModelsView,
     setAllModelsView,
     inventorySort,
@@ -1337,8 +1424,8 @@ export function ModelsPage() {
     <div className="hub-page flex min-h-0 min-w-0 flex-1 basis-0 flex-col overflow-hidden bg-background">
       <HubTopBar>
         <ModelsHeader
-          cachedCount={effectiveCachedRows.length}
-          localCount={effectiveLocalRows.length}
+          cachedCount={visibleCachedCount}
+          localCount={visibleLocalCount}
           isDataset={isDatasetMode}
           gpuLabel={gpuLabel}
           ramLabel={ramLabel}
diff --git a/studio/frontend/src/features/hub/stores/hf-token-store.ts b/studio/frontend/src/features/hub/stores/hf-token-store.ts
index 31a7f0a4da..b1e2560f02 100644
--- a/studio/frontend/src/features/hub/stores/hf-token-store.ts
+++ b/studio/frontend/src/features/hub/stores/hf-token-store.ts
@@ -116,3 +116,11 @@ export const useHfTokenStore = create<HfTokenStore>((set) => {
 export function getHfToken(): string {
   return useHfTokenStore.getState().token;
 }
+
+// HF's JS client throws on a non-empty token that isn't `hf_...` instead of
+// browsing anonymously, so treat anything malformed as no token.
+export function hfApiToken(
+  token: string | undefined | null,
+): string | undefined {
+  return token && token.startsWith("hf_") ? token : undefined;
+}
diff --git a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
index 4f275926de..7fab580f3c 100644
--- a/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
+++ b/studio/frontend/src/features/settings/stores/settings-dialog-store.ts
@@ -12,9 +12,16 @@ export type SettingsTab =
   | "api-keys"
   | "about";
 
+export type SettingsScrollTarget = "about-updates";
+
+interface OpenDialogOptions {
+  scrollTarget?: SettingsScrollTarget;
+}
+
 interface SettingsDialogState {
   open: boolean;
   activeTab: SettingsTab;
+  scrollTarget: SettingsScrollTarget | null;
   // Element focused when openDialog() ran. Radix's FocusScope normally tracks
   // this, but the rAF-scheduled focus() in settings-dialog.tsx races its
   // previous-focus capture, leaving focus on <body> after close. We restore
@@ -23,9 +30,10 @@ interface SettingsDialogState {
   // Set when something asks to jump straight to the archived chats list (the
   // archive toast). ChatTab consumes it to open the dialog, then clears it.
   archivedChatsRequested: boolean;
-  openDialog: (tab?: SettingsTab) => void;
+  openDialog: (tab?: SettingsTab, options?: OpenDialogOptions) => void;
   openArchivedChats: () => void;
   consumeArchivedChatsRequest: () => void;
+  consumeScrollTarget: (target: SettingsScrollTarget) => void;
   closeDialog: () => void;
   setActiveTab: (tab: SettingsTab) => void;
 }
@@ -65,32 +73,39 @@ function loadInitialTab(): SettingsTab {
 export const useSettingsDialogStore = create<SettingsDialogState>((set) => ({
   open: false,
   activeTab: loadInitialTab(),
+  scrollTarget: null,
   opener: null,
   archivedChatsRequested: false,
-  openDialog: (tab) =>
+  openDialog: (tab, options) =>
     set((state) => ({
       open: true,
       activeTab: tab ?? state.activeTab,
+      scrollTarget: options?.scrollTarget ?? null,
       opener: captureOpener(),
     })),
   openArchivedChats: () =>
     set({
       open: true,
       activeTab: "chat",
+      scrollTarget: null,
       archivedChatsRequested: true,
       opener: captureOpener(),
     }),
   consumeArchivedChatsRequest: () => set({ archivedChatsRequested: false }),
+  consumeScrollTarget: (target) =>
+    set((state) => ({
+      scrollTarget: state.scrollTarget === target ? null : state.scrollTarget,
+    })),
   // Do NOT clear `opener` here. onCloseAutoFocus runs on the next render
   // pass after `open: false` lands, so the opener must still be readable
   // from the store at that point. The next openDialog() overwrites it.
-  closeDialog: () => set({ open: false }),
+  closeDialog: () => set({ open: false, scrollTarget: null }),
   setActiveTab: (tab) => {
     try {
       window.localStorage.setItem(ACTIVE_TAB_KEY, tab);
     } catch {
       // ignore storage failures
     }
-    set({ activeTab: tab });
+    set({ activeTab: tab, scrollTarget: null });
   },
 }));
diff --git a/studio/frontend/src/features/settings/tabs/about-tab.tsx b/studio/frontend/src/features/settings/tabs/about-tab.tsx
index fc4a0a953e..084a3ab5b6 100644
--- a/studio/frontend/src/features/settings/tabs/about-tab.tsx
+++ b/studio/frontend/src/features/settings/tabs/about-tab.tsx
@@ -17,7 +17,7 @@ import {
   NewReleasesIcon,
 } from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
-import { useEffect, useState } from "react";
+import { useEffect, useRef, useState } from "react";
 import { SettingsRow } from "../components/settings-row";
 import { SettingsSection } from "../components/settings-section";
 import { StudioVersionSection } from "../components/studio-version-section";
@@ -25,6 +25,7 @@ import {
   type UpdateInstallSource,
   UpdateStudioInstructions,
 } from "../components/update-studio-instructions";
+import { useSettingsDialogStore } from "../stores/settings-dialog-store";
 
 type ApiObject = Record<string, unknown>;
 
@@ -76,6 +77,11 @@ export function AboutTab() {
   const deviceType = usePlatformStore((s) => s.deviceType);
   const defaultShell = deviceType === "windows" ? "windows" : "unix";
   const hw = useHardwareInfo();
+  const updateSectionRef = useRef<HTMLDivElement | null>(null);
+  const scrollTarget = useSettingsDialogStore((s) => s.scrollTarget);
+  const consumeScrollTarget = useSettingsDialogStore(
+    (s) => s.consumeScrollTarget,
+  );
   const [shutdownOpen, setShutdownOpen] = useState(false);
   const [installSource, setInstallSource] = useState<
     UpdateInstallSource | "loading"
@@ -95,6 +101,20 @@ export function AboutTab() {
     };
   }, []);
 
+  useEffect(() => {
+    if (scrollTarget !== "about-updates") {
+      return;
+    }
+    const frame = window.requestAnimationFrame(() => {
+      updateSectionRef.current?.scrollIntoView({
+        block: "start",
+        behavior: "smooth",
+      });
+      consumeScrollTarget("about-updates");
+    });
+    return () => window.cancelAnimationFrame(frame);
+  }, [consumeScrollTarget, scrollTarget]);
+
   return (
     <div className="flex flex-col gap-6">
       <header className="flex flex-col gap-1">
@@ -110,15 +130,17 @@ export function AboutTab() {
           Unsloth/Package rows; the prop keeps it About-only (General passes none). */}
       <StudioVersionSection llamaCppVersion={hw.llamaCpp} />
 
-      <SettingsSection title={t("settings.about.updates")}>
-        <div className="py-2">
-          <UpdateStudioInstructions
-            defaultShell={defaultShell}
-            installSource={isTauri ? null : installSource}
-            showTitle={false}
-          />
-        </div>
-      </SettingsSection>
+      <div ref={updateSectionRef} className="scroll-mt-5">
+        <SettingsSection title={t("settings.about.updates")}>
+          <div className="py-2">
+            <UpdateStudioInstructions
+              defaultShell={defaultShell}
+              installSource={isTauri ? null : installSource}
+              showTitle={false}
+            />
+          </div>
+        </SettingsSection>
+      </div>
 
       {hw.gpus.length > 0 || hw.cuda || hw.rocm ? (
         <SettingsSection title={t("settings.about.hardware")}>
diff --git a/studio/frontend/src/i18n/locales/en.ts b/studio/frontend/src/i18n/locales/en.ts
index ddb8286b73..7cd68865dd 100644
--- a/studio/frontend/src/i18n/locales/en.ts
+++ b/studio/frontend/src/i18n/locales/en.ts
@@ -23,6 +23,7 @@ export const en = {
     brand: "unsloth",
     product: "Unsloth Studio",
     accountMenu: "{name} account menu",
+    updateAvailable: "Update available",
     aria: {
       home: "Unsloth home",
       closeSidebar: "Close sidebar",
diff --git a/studio/frontend/src/i18n/locales/zh-CN.ts b/studio/frontend/src/i18n/locales/zh-CN.ts
index c89b1daa46..7a6a5ee66d 100644
--- a/studio/frontend/src/i18n/locales/zh-CN.ts
+++ b/studio/frontend/src/i18n/locales/zh-CN.ts
@@ -23,6 +23,7 @@ export const zhCN = {
   },
   shell: {
     accountMenu: "{name} 账号菜单",
+    updateAvailable: "有可用更新",
     aria: {
       home: "Unsloth 首页",
       closeSidebar: "关闭侧边栏",
diff --git a/studio/frontend/src/lib/bulb-icon.tsx b/studio/frontend/src/lib/bulb-icon.tsx
new file mode 100644
index 0000000000..e2f50d43b0
--- /dev/null
+++ b/studio/frontend/src/lib/bulb-icon.tsx
@@ -0,0 +1,25 @@
+// SPDX-License-Identifier: AGPL-3.0-only
+// Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+import type { ComponentPropsWithoutRef } from "react";
+
+// Shared lightbulb glyph used by the composer thinking toggle and the
+// reasoning "Thinking..." indicator so both stay in sync. Defaults are
+// overridable via props.
+export const BulbIcon = ({
+  className,
+  ...props
+}: ComponentPropsWithoutRef<"svg">) => (
+  <svg
+    className={className}
+    viewBox="-10.24 -10.24 1044.48 1044.48"
+    fill="currentColor"
+    stroke="currentColor"
+    strokeWidth={16.384}
+    xmlns="http://www.w3.org/2000/svg"
+    aria-hidden={true}
+    {...props}
+  >
+    <path d="M511.984 0c-198.032 0-353.12 161.104-353.12 359.136 0 149.2 73.28 220.256 131.185 272.128 37.28 33.424 62.368 53.552 62.368 78.352v54.255c0 1.392.193 2.752.368 4.128h-.72v92.624c.016 97.712 63.2 163.376 161.072 163.376 94.464 0 158.944-65.664 158.944-163.376V768h-.928c.176-1.376.416-2.736.416-4.128v-54.255c0-37.76 28.032-60.592 70.528-97.696 57.504-50.208 123.023-112.688 123.023-252.784C865.136 161.104 710.016 0 511.983 0zm-1.215 960c-59.904 0-94.689-37.152-94.689-99.376l-.463-42.672C438.64 825.824 470 832 512 832c41.424 0 72.848-6.624 96.08-14.768v43.392c0 63.152-35.247 99.376-97.312 99.376zm189.248-396.288c-43.472 37.968-92.433 77.216-92.433 145.904v40.432c-15.183 8.48-43.183 18.56-96.127 18.56-55.569 0-81.92-9.856-95.024-17.473V709.6c0-54.608-42.688-89.297-83.68-126.017-54.32-48.672-109.873-103.84-109.873-224.464-.015-162.72 126.385-295.12 289.104-295.12 162.752 0 289.152 132.4 289.152 295.137 0 111.024-48.463 158.576-101.12 204.576z" />
+  </svg>
+);
diff --git a/studio/install_llama_prebuilt.py b/studio/install_llama_prebuilt.py
index 1823a0f67e..33cb709ba7 100644
--- a/studio/install_llama_prebuilt.py
+++ b/studio/install_llama_prebuilt.py
@@ -230,12 +230,15 @@ FORCE_COMPILE_DEFAULT_REF = os.environ.get("UNSLOTH_LLAMA_FORCE_COMPILE_REF", "m
 _MIN_CUDA_MAJOR = 12
 _MAX_PROBE_CUDA_MAJOR = 19
 
-# Blackwell sm_120 capability thresholds. A host is Blackwell when its highest
-# compute capability is at least sm_120; ggml compiles sm_120 only at toolkit
-# >= 12.8, so an in-release windows-cuda build at or above that already covers
-# Blackwell, while cuda-12.4 does not and is dropped on a Blackwell host.
-_BLACKWELL_MIN_SM = 120
+# Blackwell floor is sm_100: data-center parts (B100/B200 sm_100, B300/GB300
+# sm_103) sit below consumer Blackwell (RTX 50 sm_120); the family needs toolkit
+# >= 12.8, except sm_103/sm_121 which need 12.9. (120 here wrongly excluded the
+# sm_100/103 data-center hosts.)
+_BLACKWELL_MIN_SM = 100
 _BLACKWELL_MIN_TOOLKIT = (12, 8)
+# SMs that need a newer toolkit than the family floor (CUDA 12.9 added native
+# sm_103/sm_121 targets; 12.8 covers sm_100/101/120).
+_BLACKWELL_SM_MIN_TOOLKIT = {103: (12, 9), 121: (12, 9)}
 
 
 def _cuda_runtime_lines_for_major(major: int) -> list[str]:
@@ -3418,18 +3421,18 @@ def windows_cuda_attempts(
     return attempts
 
 
-def _windows_cuda_attempt_covers_blackwell(attempt: AssetChoice) -> bool:
-    """True if an in-release windows-cuda attempt yields a Blackwell sm_120
-    capable build. The fork's app-named bundles declare their SM coverage
-    directly; legacy upstream-named bundles instead encode their CUDA toolkit
-    minor in the filename (covers Blackwell at toolkit >= 12.8)."""
+def _windows_cuda_attempt_covers_blackwell(
+    attempt: AssetChoice, min_toolkit: tuple[int, int] = _BLACKWELL_MIN_TOOLKIT
+) -> bool:
+    """True if a windows-cuda attempt is Blackwell-capable (app bundles via
+    declared SMs; legacy upstream bundles via toolkit minor >= min_toolkit:
+    12.8 for the family, 12.9 for sm_103/sm_121)."""
     if attempt.install_kind != "windows-cuda":
         return False
-    # Legacy upstream-named bundles encode their toolkit minor; it is the binding
-    # constraint (a 12.4 toolkit cannot offload sm_120 whatever its metadata says).
+    # Legacy bundle: the toolkit minor binds (12.4 cannot offload Blackwell).
     m = re.search(r"-bin-win-cuda-(\d+)\.(\d+)-x64\.zip$", attempt.name)
     if m is not None:
-        return (int(m.group(1)), int(m.group(2))) >= _BLACKWELL_MIN_TOOLKIT
+        return (int(m.group(1)), int(m.group(2))) >= min_toolkit
     # App-named bundles carry no minor and declare their SM coverage directly.
     return attempt.max_sm is not None and attempt.max_sm >= _BLACKWELL_MIN_SM
 
@@ -3439,22 +3442,30 @@ def _host_is_blackwell(host: HostInfo) -> bool:
     return bool(caps) and int(caps[-1]) >= _BLACKWELL_MIN_SM
 
 
+def _blackwell_min_toolkit_for_host(host: HostInfo) -> tuple[int, int]:
+    """Minimum CUDA toolkit this Blackwell host needs: 12.8 for the family,
+    12.9 if any of its SMs is sm_103/sm_121 (no native target before 12.9)."""
+    req = _BLACKWELL_MIN_TOOLKIT
+    for sm in normalize_compute_caps(host.compute_caps):
+        req = max(req, _BLACKWELL_SM_MIN_TOOLKIT.get(int(sm), _BLACKWELL_MIN_TOOLKIT))
+    return req
+
+
 def _drop_blackwell_incapable_windows_cuda(
     host: HostInfo, attempts: list[AssetChoice]
 ) -> list[AssetChoice]:
-    """On a Blackwell host, drop windows-cuda attempts that cannot offload
-    sm_120 (e.g. upstream cuda-12.4, toolkit 12.4). Such a build loads and
-    passes the functional validator but runs the model on a slow non-native
-    path (an RTX 5090 measured 7.1 tok/s vs 551.2 on cuda-13.3), so it must
-    not sit in the fallback chain behind the pin or an in-release cuda13.
-    Non-cuda attempts (windows-cpu, windows-hip, ...) pass through so the
-    host still degrades to an honest CPU install when no CUDA 13 exists."""
+    """On a Blackwell host, drop windows-cuda attempts that can't offload
+    Blackwell (e.g. cuda-12.4): they load and validate but run a slow non-native
+    path (RTX 5090: 7.1 vs 551.2 tok/s on cuda-13.3). Non-cuda attempts pass
+    through so the host can still fall back to an honest CPU install."""
     if not _host_is_blackwell(host):
         return attempts
+    min_toolkit = _blackwell_min_toolkit_for_host(host)
     return [
         attempt
         for attempt in attempts
-        if attempt.install_kind != "windows-cuda" or _windows_cuda_attempt_covers_blackwell(attempt)
+        if attempt.install_kind != "windows-cuda"
+        or _windows_cuda_attempt_covers_blackwell(attempt, min_toolkit)
     ]
 
 
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 1c3918f900..e6095748fe 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -1658,7 +1658,7 @@ def run(
 
 
 # Packages to skip on Windows (require special build steps)
-WINDOWS_SKIP_PACKAGES = {"open_spiel", "triton_kernels"}
+WINDOWS_SKIP_PACKAGES = {"triton_kernels"}
 
 # Packages to skip when torch is unavailable (Intel Mac GGUF-only mode).
 # These either *are* torch extensions or have unconditional
@@ -1675,7 +1675,6 @@ NO_TORCH_SKIP_PACKAGES = {
     "torchcodec",
     "torch-c-dlpack-ext",
     "openai-whisper",
-    "transformers-cfg",
     "librosa",
 }
 
@@ -2234,25 +2233,25 @@ def install_python_stack() -> int:
     # 4. Overrides (torchao) -- force-reinstall. The torchao version is chosen to
     #    match the torch installed in the venv so its C++ extensions load (see
     #    _select_torchao_spec). Skip when torch is unavailable (e.g. Intel Mac
-    #    GGUF-only mode): torchao requires torch.
+    #    GGUF-only mode): torchao requires torch. Also skipped on Windows ROCm
+    #    (no working build; see below).
     if NO_TORCH:
         _progress("dependency overrides (skipped, no torch)")
+    elif _rocm_windows_torch_installed:
+        # No working Windows ROCm torchao build: it imports an absent c10d backend
+        # and crashes transformers.quantizers. Studio stubs it at runtime, so
+        # installing it only ships a package that crashes on import -- skip it.
+        _progress("dependency overrides (skipped, Windows ROCm)")
+        _safe_print("   Windows ROCm -- skipping torchao (no working build; stubbed at runtime)")
     else:
         _progress("dependency overrides")
         _torch_ver = _probe_installed_torch_version()
         _torchao_spec = _select_torchao_spec(_torch_ver)
         _safe_print(f"   torch {_torch_ver or 'unknown'} detected -- installing {_torchao_spec}")
-        _override_extra_args: tuple[str, ...] = ()
-        if _rocm_windows_torch_installed:
-            # torchao declares torch as a dependency; without --no-deps uv would
-            # install CPU torch from PyPI, overwriting the AMD ROCm wheels we just
-            # installed.
-            _override_extra_args = ("--no-deps",)
         pip_install(
             "Installing dependency overrides",
             "--force-reinstall",
             "--no-cache-dir",
-            *_override_extra_args,
             _torchao_spec,
         )
 
diff --git a/studio/setup.ps1 b/studio/setup.ps1
index 20f5bd1ada..6b16702d3a 100644
--- a/studio/setup.ps1
+++ b/studio/setup.ps1
@@ -3704,35 +3704,49 @@ if (-not $NeedLlamaSourceBuild) {
         $CmakeArgs += '-DCMAKE_EXE_LINKER_FLAGS=/NODEFAULTLIB:LIBCMT'
         # CUDA flags -- only if GPU available, otherwise explicitly disable
         if ($HasNvidiaSmi -and $NvccPath) {
-            $CmakeArgs += '-DGGML_CUDA=ON'
-            # Accept a host MSVC newer than nvcc's whitelist; a fresh toolkit
-            # (e.g. CUDA 13.3) otherwise aborts with "#error -- unsupported
-            # Microsoft Visual Studio version!". Mirrors the Linux fix. Via env
-            # (covers the configure probe + build), after Refresh-Environment, idempotent.
-            $nvccAllowFlag = '-allow-unsupported-compiler'
-            if ([string]::IsNullOrEmpty($env:NVCC_PREPEND_FLAGS)) {
-                $env:NVCC_PREPEND_FLAGS = $nvccAllowFlag
-            } elseif ($env:NVCC_PREPEND_FLAGS -notlike "*$nvccAllowFlag*") {
-                $env:NVCC_PREPEND_FLAGS = "$($env:NVCC_PREPEND_FLAGS) $nvccAllowFlag"
-            }
-            substep "NVCC_PREPEND_FLAGS = $env:NVCC_PREPEND_FLAGS"
-            $CmakeArgs += "-DCUDAToolkit_ROOT=$CudaToolkitRoot"
-            $CmakeArgs += "-DCUDA_TOOLKIT_ROOT_DIR=$CudaToolkitRoot"
-            $CmakeArgs += "-DCMAKE_CUDA_COMPILER=$NvccPath"
-            if ($CudaArch) {
-                # Validate nvcc actually supports this architecture
-                if (Test-NvccArchSupport -NvccExe $NvccPath -Arch $CudaArch) {
-                    $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$CudaArch"
-                } else {
-                    # GPU arch too new for this toolkit -- fall back to highest supported.
-                    # PTX forward-compatibility will JIT-compile for the actual GPU at runtime.
-                    $maxArch = Get-NvccMaxArch -NvccExe $NvccPath
-                    if ($maxArch) {
-                        $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$maxArch"
-                        substep "GPU is sm_$CudaArch but nvcc only supports up to sm_$maxArch" "Yellow"
-                        substep "Building with sm_$maxArch (PTX will JIT for your GPU at runtime)" "Yellow"
+            # UNSLOTH_LLAMA_CUDA_ARCHS (e.g. "120" or "89;86") forces the build
+            # arch and wins over detection, matching setup.sh.
+            $CudaArchOverride = if ($env:UNSLOTH_LLAMA_CUDA_ARCHS) { ($env:UNSLOTH_LLAMA_CUDA_ARCHS -replace '\s', '') } else { '' }
+            if ((-not $CudaArch) -and (-not $CudaArchOverride)) {
+                # No detectable compute capability (#5854): -DGGML_CUDA=ON with no
+                # arch builds a PTX-only binary, so build CPU instead. Mirrors the
+                # Linux fix; set UNSLOTH_LLAMA_CUDA_ARCHS=120 to force a CUDA build.
+                substep "could not detect a CUDA compute capability; building CPU llama.cpp instead of a PTX-only binary (set UNSLOTH_LLAMA_CUDA_ARCHS=120 to force a CUDA build)." "Yellow"
+                $CmakeArgs += '-DGGML_CUDA=OFF'
+            } else {
+                $CmakeArgs += '-DGGML_CUDA=ON'
+                # Accept a host MSVC newer than nvcc's whitelist; a fresh toolkit
+                # (e.g. CUDA 13.3) otherwise aborts with "#error -- unsupported
+                # Microsoft Visual Studio version!". Mirrors the Linux fix. Via env
+                # (covers the configure probe + build), after Refresh-Environment, idempotent.
+                $nvccAllowFlag = '-allow-unsupported-compiler'
+                if ([string]::IsNullOrEmpty($env:NVCC_PREPEND_FLAGS)) {
+                    $env:NVCC_PREPEND_FLAGS = $nvccAllowFlag
+                } elseif ($env:NVCC_PREPEND_FLAGS -notlike "*$nvccAllowFlag*") {
+                    $env:NVCC_PREPEND_FLAGS = "$($env:NVCC_PREPEND_FLAGS) $nvccAllowFlag"
+                }
+                substep "NVCC_PREPEND_FLAGS = $env:NVCC_PREPEND_FLAGS"
+                $CmakeArgs += "-DCUDAToolkit_ROOT=$CudaToolkitRoot"
+                $CmakeArgs += "-DCUDA_TOOLKIT_ROOT_DIR=$CudaToolkitRoot"
+                $CmakeArgs += "-DCMAKE_CUDA_COMPILER=$NvccPath"
+                if ($CudaArchOverride) {
+                    # Forced arch wins verbatim (no nvcc validation), matching setup.sh.
+                    $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$CudaArchOverride"
+                } elseif ($CudaArch) {
+                    # Validate nvcc actually supports this architecture
+                    if (Test-NvccArchSupport -NvccExe $NvccPath -Arch $CudaArch) {
+                        $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$CudaArch"
+                    } else {
+                        # GPU arch too new for this toolkit -- fall back to highest supported.
+                        # PTX forward-compatibility will JIT-compile for the actual GPU at runtime.
+                        $maxArch = Get-NvccMaxArch -NvccExe $NvccPath
+                        if ($maxArch) {
+                            $CmakeArgs += "-DCMAKE_CUDA_ARCHITECTURES=$maxArch"
+                            substep "GPU is sm_$CudaArch but nvcc only supports up to sm_$maxArch" "Yellow"
+                            substep "Building with sm_$maxArch (PTX will JIT for your GPU at runtime)" "Yellow"
+                        }
+                        # else: omit flag entirely, let cmake pick defaults
                     }
-                    # else: omit flag entirely, let cmake pick defaults
                 }
             }
         } else {
diff --git a/studio/setup.sh b/studio/setup.sh
index ac82416f9b..e22186319f 100755
--- a/studio/setup.sh
+++ b/studio/setup.sh
@@ -155,6 +155,31 @@ _nvcc_meets_llama_minimum() {
     echo "$_raw"
 }
 
+# Echo a ';'-separated CUDA arch list (e.g. "86;120"). Override ($2,
+# UNSLOTH_LLAMA_CUDA_ARCHS) wins verbatim; else parse+dedupe compute_cap text
+# ($1). Empty means "no arch detected", so the caller builds CPU instead of a
+# PTX-only binary that fails on an old driver (#5854).
+_resolve_cuda_archs() {
+    local _raw_caps=$1
+    local _arch_override=$2
+    if [ -n "$_arch_override" ]; then
+        printf '%s' "$_arch_override"
+        return 0
+    fi
+    local _archs="" _cap _arch
+    while IFS= read -r _cap; do
+        _cap=$(printf '%s' "$_cap" | tr -d '[:space:]')
+        if [[ "$_cap" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
+            _arch="${BASH_REMATCH[1]}${BASH_REMATCH[2]}"
+            case ";$_archs;" in
+                *";$_arch;"*) ;;
+                *) _archs="${_archs:+$_archs;}$_arch" ;;
+            esac
+        fi
+    done <<< "$_raw_caps"
+    printf '%s' "$_archs"
+}
+
 # Run a GPU probe under a 10s timeout when `timeout` is available so a wedged
 # NVIDIA driver cannot hang setup; fall back to a bare call where it is not.
 _setup_run_smi() {
@@ -1517,35 +1542,38 @@ else
                     fi
 
                     if [ "$_CUDA_TOOLKIT_ALLOWED" = true ]; then
-                        CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON"
-
-                        CUDA_ARCHS=""
-                        if command -v nvidia-smi &>/dev/null; then
-                            _raw_caps=$(_setup_run_smi nvidia-smi --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)
-                            while IFS= read -r _cap; do
-                                _cap=$(echo "$_cap" | tr -d '[:space:]')
-                                if [[ "$_cap" =~ ^([0-9]+)\.([0-9]+)$ ]]; then
-                                    _arch="${BASH_REMATCH[1]}${BASH_REMATCH[2]}"
-                                    case ";$CUDA_ARCHS;" in
-                                        *";$_arch;"*) ;;
-                                        *) CUDA_ARCHS="${CUDA_ARCHS:+$CUDA_ARCHS;}$_arch" ;;
-                                    esac
-                                fi
-                            done <<< "$_raw_caps"
+                        # Resolve the arch list before committing to a CUDA build;
+                        # an empty list means CPU instead of a PTX-only binary (#5854).
+                        _raw_caps=""
+                        # Resolve nvidia-smi as _setup_has_usable_nvidia_gpu does
+                        # (PATH, then /usr/bin); `command -v` alone would miss an
+                        # off-PATH binary and wrongly drop a CUDA host to CPU.
+                        _smi_bin=""
+                        if command -v nvidia-smi >/dev/null 2>&1; then
+                            _smi_bin="nvidia-smi"
+                        elif [ -x "/usr/bin/nvidia-smi" ]; then
+                            _smi_bin="/usr/bin/nvidia-smi"
                         fi
+                        if [ -n "$_smi_bin" ]; then
+                            _raw_caps=$(_setup_run_smi "$_smi_bin" --query-gpu=compute_cap --format=csv,noheader 2>/dev/null || true)
+                        fi
+                        CUDA_ARCHS="$(_resolve_cuda_archs "$_raw_caps" "${UNSLOTH_LLAMA_CUDA_ARCHS:-}")"
 
                         if [ -n "$CUDA_ARCHS" ]; then
-                            CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
+                            CMAKE_ARGS="$CMAKE_ARGS -DGGML_CUDA=ON -DCMAKE_CUDA_ARCHITECTURES=${CUDA_ARCHS}"
+                            CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
                             _BUILD_DESC="building (CUDA, sm_${CUDA_ARCHS//;/+sm_})"
+
+                            # Allow a host gcc/clang newer than nvcc's whitelist (else a fresh
+                            # toolkit aborts with "unsupported GNU version"); via env to avoid word-splitting.
+                            export NVCC_PREPEND_FLAGS="${NVCC_PREPEND_FLAGS:+$NVCC_PREPEND_FLAGS }-allow-unsupported-compiler"
                         else
-                            _BUILD_DESC="building (CUDA)"
+                            # No detectable arch: build CPU (CMAKE_ARGS has no
+                            # -DGGML_CUDA=ON yet, so clearing GPU_BACKEND yields CPU).
+                            substep "could not detect a CUDA compute capability; building CPU llama.cpp instead of a PTX-only binary (set UNSLOTH_LLAMA_CUDA_ARCHS, e.g. \"120\", to force a CUDA build)." "$C_WARN"
+                            GPU_BACKEND=""
+                            _BUILD_DESC="building (CPU, CUDA arch undetectable)"
                         fi
-
-                        CMAKE_ARGS="$CMAKE_ARGS -DCMAKE_CUDA_FLAGS=--threads=0"
-
-                        # Allow a host gcc/clang newer than nvcc's whitelist (else a fresh
-                        # toolkit aborts with "unsupported GNU version"); via env to avoid word-splitting.
-                        export NVCC_PREPEND_FLAGS="${NVCC_PREPEND_FLAGS:+$NVCC_PREPEND_FLAGS }-allow-unsupported-compiler"
                     fi
                 fi
             elif [ "$GPU_BACKEND" = "rocm" ]; then
@@ -1794,6 +1822,7 @@ else
         printf "  ${C_DIM}%-15s${C_OK}%s${C_RST}\n" "launch" "unsloth studio -p 8888"
     fi
     printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add -H 0.0.0.0 to allow network / cloud access)"
+    printf "  ${C_DIM}%-15s%s${C_RST}\n" "" "(add --secure to allow HTTPS)"
 fi
 echo ""
 
diff --git a/studio/src-tauri/Cargo.lock b/studio/src-tauri/Cargo.lock
index c964946d0e..183cfc1cc8 100644
--- a/studio/src-tauri/Cargo.lock
+++ b/studio/src-tauri/Cargo.lock
@@ -294,6 +294,15 @@ dependencies = [
  "generic-array",
 ]
 
+[[package]]
+name = "block-buffer"
+version = "0.12.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "d2f6c7dbe95a6ed67ad9f18e57daf93a2f034c524b99fd2b76d18fdfeb6660aa"
+dependencies = [
+ "hybrid-array",
+]
+
 [[package]]
 name = "block2"
 version = "0.6.2"
@@ -518,6 +527,12 @@ dependencies = [
  "error-code",
 ]
 
+[[package]]
+name = "cmov"
+version = "0.5.4"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "0c9ea0ac24bc397ab3c98583a3c9ba74fa56b09a4449bbe172b9b1ddb016027a"
+
 [[package]]
 name = "combine"
 version = "4.6.7"
@@ -537,6 +552,12 @@ dependencies = [
  "crossbeam-utils",
 ]
 
+[[package]]
+name = "const-oid"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "a6ef517f0926dd24a1582492c791b6a4818a4d94e789a334894aa15b0d12f55c"
+
 [[package]]
 name = "convert_case"
 version = "0.4.0"
@@ -661,6 +682,15 @@ dependencies = [
  "typenum",
 ]
 
+[[package]]
+name = "crypto-common"
+version = "0.2.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "ce6e4c961d6cd6c9a86db418387425e8bdeaf05b3c8bc1411e6dca4c252f1453"
+dependencies = [
+ "hybrid-array",
+]
+
 [[package]]
 name = "cssparser"
 version = "0.29.6"
@@ -717,6 +747,15 @@ version = "0.0.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "52560adf09603e58c9a7ee1fe1dcb95a16927b17c127f0ac02d6e768a0e25bc1"
 
+[[package]]
+name = "ctutils"
+version = "0.4.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "7d5515a3834141de9eafb9717ad39eea8247b5674e6066c404e8c4b365d2a29e"
+dependencies = [
+ "cmov",
+]
+
 [[package]]
 name = "darling"
 version = "0.23.0"
@@ -823,9 +862,20 @@ version = "0.10.7"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "9ed9a281f7bc9b7576e61468ba615a66a5c8cfdff42420a70aa82701a3b1e292"
 dependencies = [
- "block-buffer",
- "crypto-common",
- "subtle",
+ "block-buffer 0.10.4",
+ "crypto-common 0.1.7",
+]
+
+[[package]]
+name = "digest"
+version = "0.11.3"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "f1dd6dbb5841937940781866fa1281a1ff7bd3bf827091440879f9994983d5c2"
+dependencies = [
+ "block-buffer 0.12.1",
+ "const-oid",
+ "crypto-common 0.2.2",
+ "ctutils",
 ]
 
 [[package]]
@@ -846,7 +896,7 @@ dependencies = [
  "libc",
  "option-ext",
  "redox_users",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -1061,7 +1111,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "39cab71617ae0d63f51a36d69f866391735b51691dbda63cf6f96d042b63efeb"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -1729,11 +1779,11 @@ checksum = "7f24254aa9a54b5c858eaee2f5bccdb46aaf0e486a595ed5fd8f86ba55232a70"
 
 [[package]]
 name = "hmac"
-version = "0.12.1"
+version = "0.13.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6c49c37c09c17a53d937dfbb742eb3a961d65a994e6bcdcf37e7399d0cc8ab5e"
+checksum = "6303bc9732ae41b04cb554b844a762b4115a61bfaa81e3e83050991eeb56863f"
 dependencies = [
- "digest",
+ "digest 0.11.3",
 ]
 
 [[package]]
@@ -1806,6 +1856,15 @@ version = "1.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "6dbf3de79e51f3d586ab4cb9d5c3e2c14aa28ed23d180cf89b4df0454a69cc87"
 
+[[package]]
+name = "hybrid-array"
+version = "0.4.12"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "9155a582abd142abc056962c29e3ce5ff2ad5469f4246b537ed42c5deba857da"
+dependencies = [
+ "typenum",
+]
+
 [[package]]
 name = "hyper"
 version = "1.8.1"
@@ -2274,9 +2333,9 @@ dependencies = [
 
 [[package]]
 name = "libc"
-version = "0.2.183"
+version = "0.2.186"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b5b646652bf6661599e1da8901b3b9522896f01e736bad5f723fe7a3a27f899d"
+checksum = "68ab91017fe16c622486840e4c83c9a37afeff978bd239b5293d61ece587de66"
 
 [[package]]
 name = "libdbus-sys"
@@ -2332,9 +2391,9 @@ dependencies = [
 
 [[package]]
 name = "log"
-version = "0.4.29"
+version = "0.4.32"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5e5032e24019045c762d3c0f28f5b6b8bbf38563a65908389bf7978758920897"
+checksum = "953f07c43838f8e6f9758cab68bf5bed85465e7587ebe0b823f1bcd81978ad3a"
 
 [[package]]
 name = "mac"
@@ -2435,9 +2494,9 @@ dependencies = [
 
 [[package]]
 name = "mio"
-version = "1.1.1"
+version = "1.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a69bcab0ad47271a0234d9422b131806bf3968021e5dc9328caf2d4cd58557fc"
+checksum = "02bd0af71c67b473010cbbc60715ee815645a4dc942899111f494b4b737d6fda"
 dependencies = [
  "libc",
  "wasi 0.11.1+wasi-snapshot-preview1",
@@ -2472,7 +2531,7 @@ dependencies = [
  "png 0.18.1",
  "serde",
  "thiserror 2.0.18",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -2902,7 +2961,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "7d8fae84b431384b68627d0f9b3b1245fcf9f46f6c0e3dc902e9dce64edd1967"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.45.0",
 ]
 
 [[package]]
@@ -3621,9 +3680,9 @@ dependencies = [
 
 [[package]]
 name = "regex"
-version = "1.12.3"
+version = "1.12.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e10754a14b9137dd7b1e3e5b0493cc9171fdd105e0ab477f51b72e7f3ac0e276"
+checksum = "f1292b7759ae1cb9ec195452d1390a074f0cd8541ab7a5a8c31cd6db45d4a6ba"
 dependencies = [
  "aho-corasick",
  "memchr",
@@ -3644,9 +3703,9 @@ dependencies = [
 
 [[package]]
 name = "regex-syntax"
-version = "0.8.10"
+version = "0.8.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc897dd8d9e8bd1ed8cdad82b5966c3e0ecae09fb1907d58efaa013543185d0a"
+checksum = "d6f6ff9a378485b298a5286656da665ba74413d36db0979633275d2e708145d4"
 
 [[package]]
 name = "reqwest"
@@ -3790,7 +3849,7 @@ dependencies = [
  "errno",
  "libc",
  "linux-raw-sys",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -3846,7 +3905,7 @@ dependencies = [
  "security-framework",
  "security-framework-sys",
  "webpki-root-certs",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -4078,9 +4137,9 @@ dependencies = [
 
 [[package]]
 name = "serde_json"
-version = "1.0.149"
+version = "1.0.150"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "83fc039473c5595ace860d8c4fafa220ff474b3fc6bfdb4293327f1a37e94d86"
+checksum = "e8014e44b4736ed0538adeecded0fce2a272f22dc9578a7eb6b2d9993c74cfb9"
 dependencies = [
  "itoa",
  "memchr",
@@ -4111,9 +4170,9 @@ dependencies = [
 
 [[package]]
 name = "serde_spanned"
-version = "1.0.4"
+version = "1.1.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f8bbf91e5a4d6315eee45e704372590b30e260ee83af6639d64557f51b067776"
+checksum = "6662b5879511e06e8999a8a235d848113e942c9124f211511b16466ee2995f26"
 dependencies = [
  "serde_core",
 ]
@@ -4210,7 +4269,18 @@ checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283"
 dependencies = [
  "cfg-if",
  "cpufeatures 0.2.17",
- "digest",
+ "digest 0.10.7",
+]
+
+[[package]]
+name = "sha2"
+version = "0.11.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "446ba717509524cb3f22f17ecc096f10f4822d76ab5c0b9822c5f9c284e825f4"
+dependencies = [
+ "cfg-if",
+ "cpufeatures 0.3.0",
+ "digest 0.11.3",
 ]
 
 [[package]]
@@ -4277,7 +4347,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "3a766e1110788c36f4fa1c2b71b387a7815aa65f88ce0229841826633d93723e"
 dependencies = [
  "libc",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -4561,9 +4631,9 @@ checksum = "61c41af27dd6d1e27b1b16b489db798443478cef1f06a660c96db617ba5de3b1"
 
 [[package]]
 name = "tauri"
-version = "2.11.1"
+version = "2.11.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b93bd86d231f0a8138f11a02a584769fe4b703dc36ae133d783228dbc4801405"
+checksum = "437404997acf375d85f1177afa7e11bb971f274ed6a7b83a2a3e339015f4cc28"
 dependencies = [
  "anyhow",
  "bytes",
@@ -4648,7 +4718,7 @@ dependencies = [
  "semver",
  "serde",
  "serde_json",
- "sha2",
+ "sha2 0.10.9",
  "syn 2.0.117",
  "tauri-utils",
  "thiserror 2.0.18",
@@ -4706,9 +4776,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-dialog"
-version = "2.7.0"
+version = "2.7.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a1fa4150c95ae391946cc8b8f905ab14797427caba3a8a2f79628e956da91809"
+checksum = "65981abb771e74e571a38196c3baa11c459379164791eba0e67abc1a5fac9884"
 dependencies = [
  "log",
  "raw-window-handle",
@@ -4724,9 +4794,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-fs"
-version = "2.5.0"
+version = "2.5.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "36e1ec28b79f3d0683f4507e1615c36292c0ea6716668770d4396b9b39871ed8"
+checksum = "b7ecc274121aca0c036a2b42d1cbe83d368d348f54e0bb8a735c2b1548e8f371"
 dependencies = [
  "anyhow",
  "dunce",
@@ -4742,7 +4812,7 @@ dependencies = [
  "tauri-plugin",
  "tauri-utils",
  "thiserror 2.0.18",
- "toml 0.9.12+spec-1.1.0",
+ "toml 1.1.2+spec-1.1.0",
  "url",
 ]
 
@@ -4767,9 +4837,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-opener"
-version = "2.5.3"
+version = "2.5.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fc624469b06f59f5a29f874bbc61a2ed737c0f9c23ef09855a292c389c42e83f"
+checksum = "17e1bea14edce6b793a04e2417e3fd924b9bc4faae83cdee7d714156cceeed29"
 dependencies = [
  "dunce",
  "glob",
@@ -4799,9 +4869,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-single-instance"
-version = "2.4.0"
+version = "2.4.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc61e4822b8f74d68278e09161d3e3fdd1b14b9eb781e24edccaabf10c420e8c"
+checksum = "5c8f29386f5e9fdc699182388a33ee80a56de436d91b67459e86afef426282af"
 dependencies = [
  "serde",
  "serde_json",
@@ -4981,10 +5051,10 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32497e9a4c7b38532efcdebeef879707aa9f794296a4f0244f6f69e9bc8574bd"
 dependencies = [
  "fastrand",
- "getrandom 0.4.2",
+ "getrandom 0.3.4",
  "once_cell",
  "rustix",
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -5116,9 +5186,9 @@ dependencies = [
 
 [[package]]
 name = "tokio"
-version = "1.50.0"
+version = "1.52.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "27ad5e34374e03cfffefc301becb44e9dc3c17584f414349ebe29ed26661822d"
+checksum = "8fc7f01b389ac15039e4dc9531aa973a135d7a4135281b12d7c1bc79fd57fffe"
 dependencies = [
  "bytes",
  "libc",
@@ -5133,9 +5203,9 @@ dependencies = [
 
 [[package]]
 name = "tokio-macros"
-version = "2.6.1"
+version = "2.7.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5c55a2eff8b69ce66c84f85e1da1c233edc36ceb85a2058d11b0d6a3c7e7569c"
+checksum = "385a6cb71ab9ab790c5fe8d67f1645e6c450a7ce006a33de03daa956cf70a496"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -5195,13 +5265,28 @@ checksum = "cf92845e79fc2e2def6a5d828f0801e29a2f8acc037becc5ab08595c7d5e9863"
 dependencies = [
  "indexmap 2.13.0",
  "serde_core",
- "serde_spanned 1.0.4",
+ "serde_spanned 1.1.1",
  "toml_datetime 0.7.5+spec-1.1.0",
  "toml_parser",
  "toml_writer",
  "winnow 0.7.15",
 ]
 
+[[package]]
+name = "toml"
+version = "1.1.2+spec-1.1.0"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "81f3d15e84cbcd896376e6730314d59fb5a87f31e4b038454184435cd57defee"
+dependencies = [
+ "indexmap 2.13.0",
+ "serde_core",
+ "serde_spanned 1.1.1",
+ "toml_datetime 1.1.1+spec-1.1.0",
+ "toml_parser",
+ "toml_writer",
+ "winnow 1.0.0",
+]
+
 [[package]]
 name = "toml_datetime"
 version = "0.6.3"
@@ -5222,9 +5307,9 @@ dependencies = [
 
 [[package]]
 name = "toml_datetime"
-version = "1.0.1+spec-1.1.0"
+version = "1.1.1+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9b320e741db58cac564e26c607d3cc1fdc4a88fd36c879568c07856ed83ff3e9"
+checksum = "3165f65f62e28e0115a00b2ebdd37eb6f3b641855f9d636d3cd4103767159ad7"
 dependencies = [
  "serde_core",
 ]
@@ -5260,25 +5345,25 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8ca1a40644a28bce036923f6a431df0b34236949d111cc07cb6dca830c9ef2e1"
 dependencies = [
  "indexmap 2.13.0",
- "toml_datetime 1.0.1+spec-1.1.0",
+ "toml_datetime 1.1.1+spec-1.1.0",
  "toml_parser",
  "winnow 1.0.0",
 ]
 
 [[package]]
 name = "toml_parser"
-version = "1.0.10+spec-1.1.0"
+version = "1.1.2+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7df25b4befd31c4816df190124375d5a20c6b6921e2cad937316de3fccd63420"
+checksum = "a2abe9b86193656635d2411dc43050282ca48aa31c2451210f4202550afb7526"
 dependencies = [
  "winnow 1.0.0",
 ]
 
 [[package]]
 name = "toml_writer"
-version = "1.0.7+spec-1.1.0"
+version = "1.1.1+spec-1.1.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f17aaa1c6e3dc22b1da4b6bba97d066e354c7945cac2f7852d4e4e7ca7a6b56d"
+checksum = "756daf9b1013ebe47a8776667b466417e2d4c5679d441c26230efd9ef78692db"
 
 [[package]]
 name = "tower"
@@ -5375,7 +5460,7 @@ dependencies = [
  "png 0.18.1",
  "serde",
  "thiserror 2.0.18",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -5403,9 +5488,9 @@ checksum = "bc7d623258602320d5c55d1bc22793b57daff0ec7efc270ea7d55ce1d5f5471c"
 
 [[package]]
 name = "typenum"
-version = "1.19.0"
+version = "1.20.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "562d481066bde0658276a35467c4af00bdc6ee726305698a55b86e61d7ad82bb"
+checksum = "b6f5e870be6c3b371b77fe0ee0bafb859fa4964b4404c27de1d380043c4dda20"
 
 [[package]]
 name = "uds_windows"
@@ -5415,7 +5500,7 @@ checksum = "f2f6fb2847f6742cd76af783a2a2c49e9375d0a111c7bef6f71cd9e738c72d6e"
 dependencies = [
  "memoffset",
  "tempfile",
- "windows-sys 0.61.2",
+ "windows-sys 0.60.2",
 ]
 
 [[package]]
@@ -5495,7 +5580,7 @@ dependencies = [
  "reqwest 0.12.28",
  "serde",
  "serde_json",
- "sha2",
+ "sha2 0.11.0",
  "simplelog",
  "tauri",
  "tauri-build",
@@ -5509,7 +5594,7 @@ dependencies = [
  "tauri-plugin-window-state",
  "tokio",
  "windows 0.62.2",
- "windows-sys 0.59.0",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -5978,7 +6063,7 @@ version = "0.1.11"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "c2a7b1c03c876122aa43f3020e6c3c3ee5c05081c9a00739faf7503aeba10d22"
 dependencies = [
- "windows-sys 0.61.2",
+ "windows-sys 0.59.0",
 ]
 
 [[package]]
@@ -6650,7 +6735,7 @@ dependencies = [
  "once_cell",
  "percent-encoding",
  "raw-window-handle",
- "sha2",
+ "sha2 0.10.9",
  "soup3",
  "tao-macros",
  "thiserror 2.0.18",
diff --git a/studio/src-tauri/Cargo.toml b/studio/src-tauri/Cargo.toml
index 2250f2774f..438558fe77 100644
--- a/studio/src-tauri/Cargo.toml
+++ b/studio/src-tauri/Cargo.toml
@@ -12,8 +12,8 @@ tauri-plugin-process = "2"
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 base64 = "0.22"
-hmac = "0.12"
-sha2 = "0.10"
+hmac = "0.13"
+sha2 = "0.11"
 reqwest = { version = "0.12", features = ["json"] }
 tokio = { version = "1", features = ["full"] }
 log = "0.4"
@@ -23,7 +23,7 @@ regex = "1"
 open = "5"
 process-wrap = { version = "9", features = ["std"] }
 fix-path-env = { git = "https://github.com/tauri-apps/fix-path-env-rs" }
-tauri-plugin-opener = "2.5.3"
+tauri-plugin-opener = "2.5.4"
 tauri-plugin-updater = "2"
 tauri-plugin-clipboard-manager = "2"
 tauri-plugin-dialog = "2"
@@ -39,7 +39,7 @@ elevated-command = "1.1.2"
 
 [target.'cfg(windows)'.dependencies]
 windows = { version = "0.62.2", features = ["Win32_System_Threading"] }
-windows-sys = { version = "0.59", features = ["Win32_Foundation", "Win32_Security", "Win32_System_Console", "Win32_System_JobObjects", "Win32_System_Threading"] }
+windows-sys = { version = "0.61", features = ["Win32_Foundation", "Win32_Security", "Win32_System_Console", "Win32_System_JobObjects", "Win32_System_Threading"] }
 
 [build-dependencies]
 tauri-build = { version = "2", features = [] }
diff --git a/studio/src-tauri/src/native_backend_lease.rs b/studio/src-tauri/src/native_backend_lease.rs
index 8c94a02f86..60c21664ec 100644
--- a/studio/src-tauri/src/native_backend_lease.rs
+++ b/studio/src-tauri/src/native_backend_lease.rs
@@ -1,5 +1,5 @@
 use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine as _};
-use hmac::{Hmac, Mac};
+use hmac::{Hmac, KeyInit, Mac};
 use serde::{Deserialize, Serialize};
 use sha2::{Digest, Sha256};
 use std::time::{Duration, SystemTime, UNIX_EPOCH};
diff --git a/tests/python/test_cpo_processor_text_tokenizer.py b/tests/python/test_cpo_processor_text_tokenizer.py
new file mode 100644
index 0000000000..69316e042d
--- /dev/null
+++ b/tests/python/test_cpo_processor_text_tokenizer.py
@@ -0,0 +1,99 @@
+"""CPO shares ORPO's row-tokenization replacements (issue #4952).
+
+CPOTrainer reuses ORPO's tokenize/init code, so the ORPO rewriters must also be
+registered for cpo_trainer. The rewriters themselves are covered by
+test_orpo_processor_text_tokenizer.py; here we just check cpo mirrors orpo.
+Static, CPU-only, no torch.
+"""
+
+import ast
+import os
+
+
+REPO_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
+RL_PATH = os.path.join(REPO_ROOT, "unsloth", "models", "rl_replacements.py")
+
+
+def _registrations(source):
+    """Map each RL_FUNCTIONS[key] target to the appended function names."""
+    out = {}
+    for node in ast.walk(ast.parse(source)):
+        if not isinstance(node, ast.Expr):
+            continue
+        call = node.value
+        if not (isinstance(call, ast.Call) and isinstance(call.func, ast.Attribute)):
+            continue
+        if call.func.attr != "append":
+            continue
+        sub = call.func.value
+        if not (isinstance(sub, ast.Subscript) and isinstance(sub.value, ast.Name)):
+            continue
+        if sub.value.id != "RL_FUNCTIONS":
+            continue
+        key = sub.slice
+        if not (isinstance(key, ast.Constant) and isinstance(key.value, str)):
+            continue
+        arg = call.args[0]
+        if isinstance(arg, ast.Name):
+            out.setdefault(key.value, []).append(arg.id)
+    return out
+
+
+def test_cpo_registration_matches_orpo():
+    regs = _registrations(open(RL_PATH).read())
+    shared = {"orpo_trainer_text_tokenizer", "orpo_trainer_processor_pad_token"}
+    assert shared <= set(regs.get("orpo_trainer", []))
+    assert shared <= set(regs.get("cpo_trainer", []))
+
+
+def _load_pad_rewriter():
+    """Exec orpo_trainer_processor_pad_token (+ _PAD_FALLBACK) without importing unsloth."""
+    tree = ast.parse(open(RL_PATH).read())
+    nodes = []
+    for n in tree.body:
+        if isinstance(n, ast.Assign) and any(
+            getattr(t, "id", None) == "_PAD_FALLBACK" for t in n.targets
+        ):
+            nodes.append(n)
+        elif isinstance(n, ast.FunctionDef) and n.name == "orpo_trainer_processor_pad_token":
+            nodes.append(n)
+    import re as _re
+
+    ns = {"re": _re}
+    exec(compile(ast.Module(body = nodes, type_ignores = []), RL_PATH, "exec"), ns)
+    return ns["orpo_trainer_processor_pad_token"]
+
+
+def test_pad_token_default_routed_through_inner_tokenizer():
+    # TRL 1.x CPO/ORPO __init__ defaults pad_token from eos_token before
+    # tokenizing; on a multimodal processor those live on `.tokenizer`. The
+    # rewrite must route both the default and pad_token_id through the inner
+    # tokenizer so a processor without bare pad_token does not AttributeError.
+    rewrite = _load_pad_rewriter()
+    init_src = (
+        "def __init__(self, model, args, processing_class):\n"
+        "    if processing_class.pad_token is None:\n"
+        "        processing_class.pad_token = processing_class.eos_token\n"
+        "    self.pad_token_id = processing_class.pad_token_id\n"
+    )
+    out = rewrite("__init__", init_src)
+    assert "if processing_class.pad_token is None:" not in out
+    assert "processing_class.pad_token = processing_class.eos_token" not in out
+    assert "_unsloth_proc_tok = getattr(processing_class, 'tokenizer', processing_class)" in out
+    # bare pad_token_id must be routed through the getattr fallback, not left raw
+    assert "= processing_class.pad_token_id\n" not in out
+    ast.parse(out)  # rewritten source still compiles
+
+
+def test_pad_rewrite_noop_without_bare_pad_block():
+    # Older TRL (the pinned <=0.24.0 range) has no bare pad_token block; the
+    # rewrite must only touch pad_token_id and leave everything else intact.
+    rewrite = _load_pad_rewriter()
+    init_src = (
+        "def __init__(self, model, args, processing_class):\n"
+        "    self.pad_token_id = processing_class.pad_token_id\n"
+    )
+    out = rewrite("__init__", init_src)
+    assert "_unsloth_proc_tok" not in out
+    assert "= processing_class.pad_token_id\n" not in out  # still routed via fallback
+    ast.parse(out)
diff --git a/tests/python/test_e2e_no_torch_sandbox.py b/tests/python/test_e2e_no_torch_sandbox.py
index 23d4a2c2b3..f412de2063 100644
--- a/tests/python/test_e2e_no_torch_sandbox.py
+++ b/tests/python/test_e2e_no_torch_sandbox.py
@@ -866,13 +866,11 @@ class TestInstallPythonStackFiltering:
 
         result_path = ips._filter_requirements(extras, ips.NO_TORCH_SKIP_PACKAGES)
         filtered = Path(result_path).read_text(encoding = "utf-8").lower()
+        lines = [
+            l.strip() for l in filtered.splitlines() if l.strip() and not l.strip().startswith("#")
+        ]
 
-        for pkg in ["torch-stoi", "timm", "openai-whisper", "transformers-cfg"]:
-            lines = [
-                l.strip()
-                for l in filtered.splitlines()
-                if l.strip() and not l.strip().startswith("#")
-            ]
+        for pkg in ips.NO_TORCH_SKIP_PACKAGES:
             assert not any(
                 l.startswith(pkg) for l in lines
             ), f"{pkg} should be removed from extras.txt"
diff --git a/tests/python/test_no_torch_filtering.py b/tests/python/test_no_torch_filtering.py
index 9dcda5c60c..732c1b7432 100644
--- a/tests/python/test_no_torch_filtering.py
+++ b/tests/python/test_no_torch_filtering.py
@@ -138,7 +138,6 @@ class TestFilterRequirements:
         req = self._write_req(
             tmp_path,
             """\
-            open_spiel
             triton_kernels
             torch-stoi
             timm
@@ -222,8 +221,8 @@ class TestRealRequirementsFiltering:
         filtered = self._non_blank_non_comment(Path(result))
         original = self._non_blank_non_comment(EXTRAS_TXT)
 
-        # These must be gone.
-        for pkg in ["torch-stoi", "timm", "openai-whisper", "transformers-cfg"]:
+        # Every NO_TORCH skip package present in extras.txt must be gone.
+        for pkg in ips.NO_TORCH_SKIP_PACKAGES:
             assert not any(
                 l.lower().startswith(pkg) for l in filtered
             ), f"{pkg} should be removed from extras.txt"
diff --git a/tests/python/test_pad_token_fix.py b/tests/python/test_pad_token_fix.py
new file mode 100644
index 0000000000..0366bee67b
--- /dev/null
+++ b/tests/python/test_pad_token_fix.py
@@ -0,0 +1,112 @@
+"""_fix_pad_token dispatch in unsloth/tokenizer_utils.py.
+
+It must delegate to unsloth_zoo's shared fix_pad_token when present (single
+source of truth), and fall back to the narrow vision-token swap against an
+older unsloth_zoo. Static + CPU-only: the two helpers are exec'd in isolation
+so the test never imports torch / transformers / unsloth.
+"""
+
+import ast
+import os
+import sys
+import types
+
+REPO_ROOT = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", ".."))
+TOK_PATH = os.path.join(REPO_ROOT, "unsloth", "tokenizer_utils.py")
+
+WANTED = {
+    "_VISION_PAD_TOKENS",
+    "_SAFE_TEXT_PAD_TOKENS",
+    "_fix_vision_pad_token",
+    "_fix_pad_token",
+}
+
+
+def _load_pad_helpers():
+    """Exec only the pad-token helpers with a stub logger (no heavy imports)."""
+    tree = ast.parse(open(TOK_PATH).read())
+    nodes = []
+    for node in tree.body:
+        if isinstance(node, ast.Assign):
+            names = {t.id for t in node.targets if isinstance(t, ast.Name)}
+            if names & WANTED:
+                nodes.append(node)
+        elif isinstance(node, ast.FunctionDef) and node.name in WANTED:
+            nodes.append(node)
+    module = ast.Module(body = nodes, type_ignores = [])
+    ast.fix_missing_locations(module)
+    ns = {"logger": types.SimpleNamespace(warning = lambda *a, **k: None)}
+    exec(compile(module, TOK_PATH, "exec"), ns)
+    return ns
+
+
+class FakeTok:
+    def __init__(self, vocab, pad, eos):
+        self._v = dict(vocab)
+        self.pad_token = pad
+        self.eos_token = eos
+
+    @property
+    def pad_token_id(self):
+        return self._v.get(self.pad_token)
+
+    @property
+    def eos_token_id(self):
+        return self._v.get(self.eos_token)
+
+    def get_vocab(self):
+        return dict(self._v)
+
+
+def _block_shared_module(monkeypatch):
+    # Stub parent so importing it is cheap, and mark the submodule absent.
+    monkeypatch.setitem(sys.modules, "unsloth_zoo", types.ModuleType("unsloth_zoo"))
+    monkeypatch.setitem(sys.modules, "unsloth_zoo.pad_token", None)
+
+
+def test_fix_pad_token_none_is_noop():
+    ns = _load_pad_helpers()
+    assert ns["_fix_pad_token"](None) is None
+
+
+def test_fix_pad_token_falls_back_without_shared_module(monkeypatch):
+    ns = _load_pad_helpers()
+    _block_shared_module(monkeypatch)
+    # Qwen3 text tokenizer shipping a vision pad_token -> narrow swap heals it.
+    tok = FakeTok(
+        {"<|endoftext|>": 1, "<|im_end|>": 2, "<|vision_pad|>": 3},
+        pad = "<|vision_pad|>",
+        eos = "<|im_end|>",
+    )
+    ns["_fix_pad_token"](tok)
+    assert tok.pad_token == "<|endoftext|>"
+    assert tok.pad_token != tok.eos_token
+
+
+def test_fallback_leaves_valid_pad_untouched(monkeypatch):
+    ns = _load_pad_helpers()
+    _block_shared_module(monkeypatch)
+    tok = FakeTok({"<s>": 1, "</s>": 2, "<pad>": 0}, pad = "<pad>", eos = "</s>")
+    ns["_fix_pad_token"](tok)
+    assert tok.pad_token == "<pad>"
+
+
+def test_fix_pad_token_delegates_to_shared_module(monkeypatch):
+    ns = _load_pad_helpers()
+    calls = {}
+
+    fake = types.ModuleType("unsloth_zoo.pad_token")
+
+    def fix_pad_token(tokenizer, allow_add = True):
+        calls["allow_add"] = allow_add
+        tokenizer.pad_token = "SHARED"
+
+    fake.fix_pad_token = fix_pad_token
+    monkeypatch.setitem(sys.modules, "unsloth_zoo", types.ModuleType("unsloth_zoo"))
+    monkeypatch.setitem(sys.modules, "unsloth_zoo.pad_token", fake)
+
+    tok = FakeTok({"a": 1}, pad = "x", eos = "y")
+    ns["_fix_pad_token"](tok)
+    # Delegated, and crucially with allow_add=False (no model here to resize).
+    assert tok.pad_token == "SHARED"
+    assert calls["allow_add"] is False
diff --git a/tests/run_all.sh b/tests/run_all.sh
index 2c495f5561..18182d9db7 100755
--- a/tests/run_all.sh
+++ b/tests/run_all.sh
@@ -11,6 +11,7 @@ sh "$TESTS_DIR/sh/test_get_torch_index_url.sh"
 sh "$TESTS_DIR/sh/test_mac_intel_compat.sh"
 sh "$TESTS_DIR/sh/test_torch_constraint.sh"
 sh "$TESTS_DIR/sh/test_nvcc_meets_llama_minimum.sh"
+sh "$TESTS_DIR/sh/test_resolve_cuda_archs.sh"
 sh "$TESTS_DIR/sh/test_strixhalo_wsl_reroute.sh"
 sh "$TESTS_DIR/sh/test_uninstall_shared_icon.sh"
 sh "$TESTS_DIR/sh/test_torch_flavor.sh"
diff --git a/tests/sh/test_resolve_cuda_archs.sh b/tests/sh/test_resolve_cuda_archs.sh
new file mode 100755
index 0000000000..be32bee6ad
--- /dev/null
+++ b/tests/sh/test_resolve_cuda_archs.sh
@@ -0,0 +1,68 @@
+#!/bin/bash
+# Unit tests for _resolve_cuda_archs() from studio/setup.sh (#5854).
+# Turns nvidia-smi compute_cap text into a deduped ';'-separated arch list;
+# empty result signals a CPU build, and an explicit override wins.
+set -e
+
+SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
+SETUP_SH="$SCRIPT_DIR/../../studio/setup.sh"
+PASS=0
+FAIL=0
+
+# Extract just the helper function (same sed range as the other function tests).
+_FUNC_FILE=$(mktemp)
+sed -n '/^_resolve_cuda_archs()/,/^}/p' "$SETUP_SH" > "$_FUNC_FILE"
+
+assert_eq() {
+    _label="$1"; _expected="$2"; _actual="$3"
+    if [ "$_actual" = "$_expected" ]; then
+        echo "  PASS: $_label"
+        PASS=$((PASS + 1))
+    else
+        echo "  FAIL: $_label (expected '$_expected', got '$_actual')"
+        FAIL=$((FAIL + 1))
+    fi
+}
+
+# $1 = raw compute_cap text, $2 = override
+run_resolve() {
+    bash -c ". '$_FUNC_FILE'; _resolve_cuda_archs \"\$1\" \"\$2\"" _ "$1" "$2"
+}
+
+echo "=== test_resolve_cuda_archs ==="
+
+# 1) Single GPU -> single arch.
+assert_eq "single 8.6" "86" "$(run_resolve "8.6" "")"
+
+# 2) Two distinct GPUs -> both archs, order preserved.
+assert_eq "distinct 8.6 + 9.0" "86;90" "$(run_resolve "$(printf '8.6\n9.0\n')" "")"
+
+# 3) Duplicate caps (multi-GPU same model) -> deduped.
+assert_eq "dedup 12.0 x2" "120" "$(run_resolve "$(printf '12.0\n12.0\n')" "")"
+
+# 4) Empty input -> empty (the CPU-fallback signal; #5854).
+assert_eq "empty input" "" "$(run_resolve "" "")"
+
+# 5) Garbage / N/A lines are ignored -> empty.
+assert_eq "garbage N/A" "" "$(run_resolve "$(printf 'N/A\n[Not Supported]\n')" "")"
+
+# 6) Mixed valid + junk -> only the valid caps survive.
+assert_eq "mixed valid+junk" "86;90" "$(run_resolve "$(printf '8.6\nfoo\n9.0\n')" "")"
+
+# 7) Whitespace / CR around a cap is stripped.
+assert_eq "whitespace stripped" "86" "$(run_resolve "$(printf '  8.6 \r\n')" "")"
+
+# 8) Override wins verbatim, ignoring detection.
+assert_eq "override wins" "120" "$(run_resolve "8.6" "120")"
+
+# 9) Override works even with no detected caps.
+assert_eq "override no detection" "86;90" "$(run_resolve "" "86;90")"
+
+# 10) Future arch (compute 10.0 -> 100) parses.
+assert_eq "future 10.0" "100" "$(run_resolve "10.0" "")"
+
+rm -f "$_FUNC_FILE"
+
+echo ""
+echo "Results: $PASS passed, $FAIL failed"
+[ "$FAIL" -eq 0 ] || exit 1
diff --git a/tests/studio/install/test_gpu_detection_followups.py b/tests/studio/install/test_gpu_detection_followups.py
index 0926cf9bd5..f5ad9566d7 100644
--- a/tests/studio/install/test_gpu_detection_followups.py
+++ b/tests/studio/install/test_gpu_detection_followups.py
@@ -249,7 +249,20 @@ class TestSetupShHardening:
         ), "ROCm toolkit search must require a detected AMD GPU, not just hipcc"
 
     def test_compute_cap_probe_timeout_wrapped(self, setup_src):
-        assert "_setup_run_smi nvidia-smi --query-gpu=compute_cap" in setup_src
+        # nvidia-smi is now a variable ($_smi_bin), so check the wrapper precedes
+        # the probe rather than matching a literal. The string also appears in a
+        # comment, so scan all occurrences and accept if any is wrapped.
+        wrapped = False
+        start = 0
+        while True:
+            idx = setup_src.find("--query-gpu=compute_cap", start)
+            if idx < 0:
+                break
+            if "_setup_run_smi" in setup_src[max(0, idx - 80) : idx]:
+                wrapped = True
+                break
+            start = idx + 1
+        assert wrapped, "compute_cap probe must be wrapped in _setup_run_smi (timeout-bounded)"
 
     def test_driver_version_probe_timeout_wrapped(self, setup_src):
         start = setup_src.find("_cuda_driver_max_version()")
diff --git a/tests/studio/install/test_pr4562_bugfixes.py b/tests/studio/install/test_pr4562_bugfixes.py
index b19c300fec..a8cec0cb85 100644
--- a/tests/studio/install/test_pr4562_bugfixes.py
+++ b/tests/studio/install/test_pr4562_bugfixes.py
@@ -728,13 +728,16 @@ class TestSourceCodePatterns:
         assert all(
             "-allow-unsupported-compiler" not in line for line in cmake_args_lines
         ), "flag must not be pushed into the $CmakeArgs array"
-        # Must be scoped to the CUDA branch, not set for CPU-only builds.
+        # Must be scoped to the CUDA-on branch, not set for CPU-only builds. The
+        # branch also has an early GGML_CUDA=OFF (undetectable-arch CPU fallback,
+        # #5854), so anchor on GGML_CUDA=ON and the final (no-GPU) GGML_CUDA=OFF.
         flag_idx = content.index("-allow-unsupported-compiler")
         cuda_guard_idx = content.index("if ($HasNvidiaSmi -and $NvccPath)")
-        cuda_disable_idx = content.index("'-DGGML_CUDA=OFF'")
-        assert cuda_guard_idx < flag_idx < cuda_disable_idx, (
-            "NVCC_PREPEND_FLAGS must be set inside the CUDA-on branch, "
-            "before the GGML_CUDA=OFF (CPU) branch"
+        cuda_on_idx = content.index("'-DGGML_CUDA=ON'")
+        cpu_else_idx = content.rindex("'-DGGML_CUDA=OFF'")
+        assert cuda_guard_idx < cuda_on_idx < flag_idx < cpu_else_idx, (
+            "NVCC_PREPEND_FLAGS must be set inside the CUDA-on branch, after "
+            "-DGGML_CUDA=ON and before the final CPU GGML_CUDA=OFF branch"
         )
 
     def test_macos_arm64_cpu_fallback_args_exclude_rpath(self):
diff --git a/tests/studio/install/test_rocm_support.py b/tests/studio/install/test_rocm_support.py
index 6915269cec..4c50fa1c8e 100644
--- a/tests/studio/install/test_rocm_support.py
+++ b/tests/studio/install/test_rocm_support.py
@@ -1221,6 +1221,8 @@ _WORKER_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "training" / "work
 _EXPORT_WORKER_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "export" / "worker.py"
 # Shared torchao Windows-ROCm stub used by both workers.
 _TORCHAO_STUB_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "_torchao_stub.py"
+# RAG embedder -- runs in the main backend process and also needs the stub.
+_EMBEDDINGS_PATH = PACKAGE_ROOT / "studio" / "backend" / "core" / "rag" / "embeddings.py"
 # Wheel-probe script literal lives in wheel_utils after the resolver refactor.
 _WHEEL_UTILS_PATH = PACKAGE_ROOT / "studio" / "backend" / "utils" / "wheel_utils.py"
 
@@ -2329,6 +2331,13 @@ class TestWorkerWindowsRocmPatches:
         source = _EXPORT_WORKER_PATH.read_text(encoding = "utf-8")
         assert "install_torchao_windows_rocm_stub()" in source
 
+    def test_embedder_calls_shared_torchao_stub(self):
+        """embeddings.py must install the stub before importing sentence-transformers:
+        it runs in the main process (not a stubbed worker), so otherwise transformers
+        -> torchao crashes on Windows ROCm and the embedder drops to llama-server."""
+        source = _EMBEDDINGS_PATH.read_text(encoding = "utf-8")
+        assert "install_torchao_windows_rocm_stub()" in source
+
     def test_torchao_stub_uses_stub_type_meta(self):
         """Torchao stub must use _StubTypeMeta so isinstance() returns False not TypeError."""
         source = _TORCHAO_STUB_PATH.read_text(encoding = "utf-8")
diff --git a/tests/studio/test_stream_cancel_registration_timing.py b/tests/studio/test_stream_cancel_registration_timing.py
index 2e6b5f14da..33deb7af9d 100644
--- a/tests/studio/test_stream_cancel_registration_timing.py
+++ b/tests/studio/test_stream_cancel_registration_timing.py
@@ -146,24 +146,50 @@ def test_async_generators_cleanup_tracker_in_finally():
     )
 
 
-def test_streaming_responses_have_no_background_task():
-    top = None
-    for n in ast.walk(_TREE):
-        if isinstance(n, ast.AsyncFunctionDef) and n.name == "openai_chat_completions":
-            top = n
-            break
-    assert top is not None
+def test_chat_completions_streams_avoid_starlette_task_group():
+    top = _async_function("openai_chat_completions")
+    legacy_calls = []
+    same_task_calls = 0
     for sub in ast.walk(top):
         if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
             continue
-        if sub.func.id != "StreamingResponse":
+        if sub.func.id == "StreamingResponse":
+            legacy_calls.append(sub.lineno)
+        if sub.func.id == "_SameTaskStreamingResponse":
+            same_task_calls += 1
+    assert not legacy_calls, (
+        "Streaming /v1/chat/completions must use _SameTaskStreamingResponse, "
+        "not Starlette's legacy task-group StreamingResponse. Lines: "
+        f"{legacy_calls}"
+    )
+    assert same_task_calls >= 5
+
+
+def test_openai_passthrough_stream_avoids_starlette_task_group():
+    top = _async_function("_openai_passthrough_stream")
+    legacy_calls = []
+    same_task_calls = 0
+    for sub in ast.walk(top):
+        if not (isinstance(sub, ast.Call) and isinstance(sub.func, ast.Name)):
             continue
-        kwargs = {kw.arg for kw in sub.keywords if kw.arg}
-        assert "background" not in kwargs, (
-            "StreamingResponse in openai_chat_completions must not pass "
-            "`background=` -- cleanup now lives in the generator's finally "
-            "block; a BackgroundTask would be skipped on abrupt disconnect"
-        )
+        if sub.func.id == "StreamingResponse":
+            legacy_calls.append(sub.lineno)
+        if sub.func.id == "_SameTaskStreamingResponse":
+            same_task_calls += 1
+    assert not legacy_calls, (
+        "OpenAI passthrough streams must use _SameTaskStreamingResponse, "
+        "not Starlette's legacy task-group StreamingResponse. Lines: "
+        f"{legacy_calls}"
+    )
+    assert same_task_calls >= 2
+
+
+def test_local_chat_streams_install_same_task_disconnect_watcher():
+    top = _async_function("openai_chat_completions")
+    assert _calls_name(top, "_await_disconnect_then_cancel"), (
+        "Local same-task streams must watch request disconnects themselves; "
+        "do not restore Starlette's task-group StreamingResponse for this."
+    )
 
 
 def test_direct_llama_server_streams_install_disconnect_watcher():
@@ -185,6 +211,33 @@ def test_direct_llama_server_streams_install_disconnect_watcher():
     )
 
 
+def test_audio_input_stream_installs_disconnect_watcher():
+    audio = _async_function("audio_input_stream")
+    has_watcher = False
+    has_cleanup = False
+    for sub in ast.walk(audio):
+        if isinstance(sub, ast.Call):
+            fn = sub.func
+            if (
+                isinstance(fn, ast.Attribute)
+                and fn.attr == "create_task"
+                and isinstance(fn.value, ast.Name)
+                and fn.value.id == "asyncio"
+                and sub.args
+                and isinstance(sub.args[0], ast.Call)
+                and isinstance(sub.args[0].func, ast.Name)
+                and sub.args[0].func.id == "_await_disconnect_then_cancel"
+            ):
+                has_watcher = True
+            if isinstance(fn, ast.Name) and fn.id == "_stop_local_disconnect_cancel_watcher":
+                has_cleanup = True
+    assert has_watcher, (
+        "audio_input_stream must install a disconnect watcher so client "
+        "disconnects set cancel_event while asyncio.to_thread(next, ...) is blocked"
+    )
+    assert has_cleanup, "audio_input_stream must stop its disconnect watcher in finally"
+
+
 # ── Behavioral helpers ───────────────────────────────────────
 
 _WANTED = {
@@ -222,6 +275,21 @@ def _load_registry_module():
     return mod
 
 
+def _load_same_task_response_module():
+    for n in _TREE.body:
+        if isinstance(n, ast.ClassDef) and n.name == "_SameTaskStreamingResponse":
+            source = ast.get_source_segment(SRC, n)
+            break
+    else:
+        raise AssertionError("_SameTaskStreamingResponse missing")
+    mod = {}
+    exec(
+        "class StreamingResponse: pass\nclass ClientDisconnect(Exception): pass\n" + source,
+        mod,
+    )
+    return mod
+
+
 def _make_stream(tracker, raise_exc):
     async def gen():
         try:
@@ -326,6 +394,39 @@ def test_finally_cleanup_on_aclose():
     assert "sid-abort" not in m["_CANCEL_REGISTRY"]
 
 
+def test_same_task_response_closes_body_iterator_on_send_disconnect():
+    m = _load_same_task_response_module()
+    closed = False
+
+    async def body():
+        nonlocal closed
+        try:
+            yield "data: first\n\n"
+        finally:
+            closed = True
+
+    async def run():
+        agen = body()
+        await agen.__anext__()
+        response = m["_SameTaskStreamingResponse"].__new__(m["_SameTaskStreamingResponse"])
+        response.body_iterator = agen
+        response.background = None
+
+        async def stream_response(_send):
+            raise OSError("client disconnected")
+
+        response.stream_response = stream_response
+        try:
+            await response({}, None, lambda _message: None)
+        except m["ClientDisconnect"]:
+            pass
+        else:
+            raise AssertionError("expected ClientDisconnect")
+
+    asyncio.run(run())
+    assert closed
+
+
 def test_preset_cancel_event_exits_cleanly_with_done():
     # Pending-replay: a stashed cancel pre-set cancel_event. The loop must break
     # cleanly with final_chunk + [DONE], not propagate GeneratorExit from the GGUF wrapper.
diff --git a/tests/test_fused_ce_not_return_dict_logits.py b/tests/test_fused_ce_not_return_dict_logits.py
new file mode 100644
index 0000000000..51a6ec4843
--- /dev/null
+++ b/tests/test_fused_ce_not_return_dict_logits.py
@@ -0,0 +1,62 @@
+# Unsloth - 2x faster, 60% less VRAM LLM training and finetuning
+# Copyright 2023-present Daniel Han-Chen, Michael Han-Chen & the Unsloth team. All rights reserved.
+#
+# This program is free software: you can redistribute it and/or modify
+# it under the terms of the GNU Lesser General Public License as published by
+# the Free Software Foundation, either version 3 of the License, or
+# (at your option) any later version.
+#
+# This program is distributed in the hope that it will be useful,
+# but WITHOUT ANY WARRANTY; without even the implied warranty of
+# MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
+# GNU Lesser General Public License for more details.
+
+"""Drift detector for unsloth#2068.
+
+The fused-CE path never materializes ``logits``, so its ``not return_dict``
+return must use ``EMPTY_LOGITS`` (it once used ``logits`` -> UnboundLocalError).
+Pure text inspection, so it runs GPU-free (the fused path is GPU/triton only)."""
+
+from __future__ import annotations
+
+import re
+from pathlib import Path
+
+import pytest
+
+
+_REPO = Path(__file__).resolve().parent.parent
+
+
+def _fused_ce_not_return_dict_return(source: str) -> str:
+    """Return the `output = (...) + outputs[1:]` assignment from the first
+    `if not return_dict:` after the `unsloth_fused_ce_loss(` call. Whitespace-
+    tolerant regexes survive reformatting; the anchor targets the fused branch."""
+    fused = re.search(r"unsloth_fused_ce_loss\s*\(", source)
+    if fused is None:
+        raise ValueError("unsloth_fused_ce_loss( call not found")
+    guard = re.search(r"if\s+not\s+return_dict\s*:", source[fused.end() :])
+    if guard is None:
+        raise ValueError("`if not return_dict:` guard not found after fused-CE call")
+    guard_start = fused.end() + guard.start()
+    out = re.search(r"output\s*=\s*\(.*?outputs\s*\[\s*1\s*:\s*\]", source[guard_start:], re.DOTALL)
+    if out is None:
+        raise ValueError("`output = (...) + outputs[1:]` assignment not found")
+    return out.group(0)
+
+
+@pytest.mark.parametrize("rel", ["unsloth/models/llama.py", "unsloth/models/mistral.py"])
+def test_fused_ce_not_return_dict_uses_empty_logits(rel):
+    path = _REPO / rel
+    source = path.read_text(encoding = "utf-8")
+    assert "unsloth_fused_ce_loss(" in source, f"{rel}: fused-CE call vanished"
+
+    ret = _fused_ce_not_return_dict_return(source)
+    assert "EMPTY_LOGITS" in ret, (
+        f"DRIFT (#2068): {rel} fused-CE `not return_dict` returns {ret!r}; it must "
+        f"use EMPTY_LOGITS since `logits` is never assigned on the fused-CE path."
+    )
+    assert "(logits,)" not in ret, (
+        f"DRIFT (#2068): {rel} fused-CE `not return_dict` references the unassigned "
+        f"`logits` ({ret!r}) -> UnboundLocalError at runtime."
+    )
diff --git a/tests/test_missing_torchvision_vlm.py b/tests/test_missing_torchvision_vlm.py
new file mode 100644
index 0000000000..5bc72a21da
--- /dev/null
+++ b/tests/test_missing_torchvision_vlm.py
@@ -0,0 +1,30 @@
+"""Regression test for unsloth#4202: detect missing torchvision so the loader can
+surface the real cause instead of a misleading collator error."""
+
+import importlib.util
+from unittest import mock
+
+from unsloth.models.vision import _missing_torchvision_error
+
+
+def test_error_text_mentions_torchvision_is_detected():
+    err = ImportError("Qwen3VLVideoProcessor requires the Torchvision library but ...")
+    assert _missing_torchvision_error(err) is True
+
+
+def test_torchvision_missing_is_detected_without_error():
+    with mock.patch.object(importlib.util, "find_spec", return_value = None):
+        assert _missing_torchvision_error(None) is True
+
+
+def test_torchvision_present_unrelated_error_is_not_flagged():
+    sentinel = object()
+    with mock.patch.object(importlib.util, "find_spec", return_value = sentinel):
+        assert _missing_torchvision_error(ValueError("unrelated")) is False
+        assert _missing_torchvision_error(None) is False
+
+
+def test_matches_real_environment():
+    # find_spec is the source of truth when no error is supplied.
+    expected = importlib.util.find_spec("torchvision") is None
+    assert _missing_torchvision_error(None) is expected
diff --git a/tests/utils/test_attention_dispatch_dora_dtype.py b/tests/utils/test_attention_dispatch_dora_dtype.py
new file mode 100644
index 0000000000..a6586dea8d
--- /dev/null
+++ b/tests/utils/test_attention_dispatch_dora_dtype.py
@@ -0,0 +1,73 @@
+"""Regression test for unslothai/unsloth#1013: run_attention must downcast DoRA's
+fp32 q/k/v for FlashAttention while leaving already-16bit tensors untouched."""
+
+import torch
+import unsloth  # noqa: F401
+
+from unsloth.utils import attention_dispatch as ad
+
+
+def _run(monkeypatch, qkv_dtype, backend):
+    captured = {}
+
+    def _check(Q):
+        captured["dtype"] = Q.dtype
+        # Mirror the real kernel constraint so an unfixed dispatch fails loudly.
+        if Q.dtype not in (torch.float16, torch.bfloat16):
+            raise RuntimeError("FlashAttention only support fp16 and bf16 data type")
+
+    def fake_flash_dense(Q, K, V, **kwargs):
+        _check(Q)
+        return torch.zeros_like(Q)
+
+    def fake_flash_varlen(Q, K, V, *args, **kwargs):
+        _check(Q)
+        return torch.zeros_like(Q)
+
+    monkeypatch.setattr(ad, "flash_attn_func", fake_flash_dense, raising = False)
+    monkeypatch.setattr(ad, "flash_attn_varlen_func", fake_flash_varlen, raising = False)
+
+    bsz, n_heads, q_len, head_dim = 1, 2, 4, 8
+    Q = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+    K = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+    V = torch.randn(bsz, n_heads, q_len, head_dim, dtype = qkv_dtype)
+
+    seq_info = None
+    if backend == ad.FLASH_VARLEN:
+        cu = torch.tensor([0, q_len], dtype = torch.int32)
+        seq_info = (None, cu, q_len)
+
+    config = ad.AttentionConfig(
+        backend = backend,
+        n_kv_heads = n_heads,
+        n_groups = 1,
+        flash_dense_kwargs = {"causal": True},
+        flash_varlen_kwargs = {"dropout_p": 0.0, "causal": True},
+    )
+    context = ad.AttentionContext(
+        bsz = bsz,
+        q_len = q_len,
+        kv_seq_len = q_len,
+        n_heads = n_heads,
+        head_dim = head_dim,
+        requires_grad = False,
+        seq_info = seq_info,
+        attention_mask = None,
+        causal_mask = None,
+    )
+    ad.run_attention(config = config, context = context, Q = Q, K = K, V = V)
+    return captured["dtype"]
+
+
+def test_dense_flash_downcasts_fp32_qkv(monkeypatch):
+    # fp32 DoRA output must be downcast to a flash-compatible dtype.
+    assert _run(monkeypatch, torch.float32, ad.FLASH_DENSE) in (torch.bfloat16, torch.float16)
+
+
+def test_varlen_flash_downcasts_fp32_qkv(monkeypatch):
+    assert _run(monkeypatch, torch.float32, ad.FLASH_VARLEN) in (torch.bfloat16, torch.float16)
+
+
+def test_bf16_qkv_left_untouched(monkeypatch):
+    # Standard LoRA path (already bf16) must not be altered.
+    assert _run(monkeypatch, torch.bfloat16, ad.FLASH_DENSE) == torch.bfloat16
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index 2365975cdd..7a056cef82 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.6.8"
+__version__ = "2026.6.9"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",
diff --git a/unsloth/models/llama.py b/unsloth/models/llama.py
index 7b9cf3df0e..e8584caf7d 100644
--- a/unsloth/models/llama.py
+++ b/unsloth/models/llama.py
@@ -1486,7 +1486,9 @@ def CausalLM_fast_forward(fast_forward_inference):
                     logit_softcapping = logit_softcapping,
                 )
                 if not return_dict:
-                    output = (logits,) + outputs[1:]
+                    # Fused CE never materializes `logits`; use EMPTY_LOGITS
+                    # like the return_dict branch below (fixes #2068).
+                    output = (EMPTY_LOGITS,) + outputs[1:]
                     return (loss,) + output if loss is not None else output
 
                 output = CausalLMOutputWithPast(
@@ -2931,6 +2933,7 @@ class FastLlamaModel:
                 ("finetune_language_layers", True),
                 ("finetune_attention_modules", True),
                 ("finetune_mlp_modules", True),
+                ("finetune_audio_layers", False),
             ):
                 if peft_arg not in kwargs:
                     kwargs[peft_arg] = flag
diff --git a/unsloth/models/mistral.py b/unsloth/models/mistral.py
index 74f7acbb03..df2a4de5bd 100644
--- a/unsloth/models/mistral.py
+++ b/unsloth/models/mistral.py
@@ -332,7 +332,9 @@ def MistralForCausalLM_fast_forward(
                 logit_softcapping = logit_softcapping,
             )
             if not return_dict:
-                output = (logits,) + outputs[1:]
+                # Fused CE never materializes `logits`; use EMPTY_LOGITS
+                # like the return_dict branch below (fixes #2068).
+                output = (EMPTY_LOGITS,) + outputs[1:]
                 return (loss,) + output if loss is not None else output
 
             output = CausalLMOutputWithPast(
diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py
index ab18a5ec93..61a07b686d 100644
--- a/unsloth/models/rl_replacements.py
+++ b/unsloth/models/rl_replacements.py
@@ -511,8 +511,8 @@ def sft_trainer_compute_loss(function_name, function):
 RL_FUNCTIONS["sft_trainer"].append(sft_trainer_compute_loss)
 
 
-# Use the underlying text tokenizer for ORPO row tokenization when a
-# multimodal processor is supplied as the processing class.
+# Route ORPO/CPO row tokenization through the underlying text tokenizer when the
+# processing class is a multimodal processor; CPO reuses this code (#4952).
 def orpo_trainer_text_tokenizer(function_name, function):
     if function_name == "build_tokenized_answer":
         function = re.sub(
@@ -555,11 +555,12 @@ def orpo_trainer_text_tokenizer(function_name, function):
 
 
 RL_FUNCTIONS["orpo_trainer"].append(orpo_trainer_text_tokenizer)
+RL_FUNCTIONS["cpo_trainer"].append(orpo_trainer_text_tokenizer)
 
 
 # Resolve `processing_class.pad_token_id` through the underlying tokenizer when
 # a multimodal processor is supplied (processors lack `pad_token_id`). Without
-# this, ORPOTrainer.__init__ raises AttributeError on
+# this, ORPO/CPOTrainer.__init__ raises AttributeError on
 # `DPODataCollatorWithPadding(pad_token_id=processing_class.pad_token_id, ...)`
 # and on `self.padding_value = ... else processing_class.pad_token_id`.
 _PAD_FALLBACK = (
@@ -572,12 +573,28 @@ _PAD_FALLBACK = (
 def orpo_trainer_processor_pad_token(function_name, function):
     if function_name != "__init__":
         return function
+    # Multimodal processors (e.g. Gemma3/Gemma4 Processor) expose pad_token /
+    # eos_token on `.tokenizer`, not on the processor itself. TRL 1.x CPO/ORPO
+    # __init__ defaults `processing_class.pad_token` from `.eos_token` before
+    # tokenizing, which AttributeErrors on such a processor. Route the default
+    # through the inner tokenizer. Older TRL lacks this block, so the sub is a
+    # no-op there and only the pad_token_id fallback below applies.
+    function = re.sub(
+        r"(?m)^([ \t]*)if processing_class\.pad_token is None:\n"
+        r"\1[ \t]+processing_class\.pad_token\s*=\s*processing_class\.eos_token\n",
+        r"\1_unsloth_proc_tok = getattr(processing_class, 'tokenizer', processing_class)\n"
+        r"\1if getattr(_unsloth_proc_tok, 'pad_token', None) is None:\n"
+        r"\1    _unsloth_proc_tok.pad_token = getattr(_unsloth_proc_tok, 'eos_token', None)\n",
+        function,
+        count = 1,
+    )
     if "processing_class.pad_token_id" not in function:
         return function
     return function.replace("processing_class.pad_token_id", _PAD_FALLBACK)
 
 
 RL_FUNCTIONS["orpo_trainer"].append(orpo_trainer_processor_pad_token)
+RL_FUNCTIONS["cpo_trainer"].append(orpo_trainer_processor_pad_token)
 
 
 # Fix bare pop("push_to_hub_token") in compiled SFT/IterativeSFT trainer __init__
@@ -1530,7 +1547,11 @@ def grpo_trainer_compute_loss(function_name, function):
         num_items_in_batch = inputs.get("num_items_in_batch", None)
         sampling_per_token_logps = inputs.get("sampling_per_token_logps", None)
         tool_mask = inputs.get("tool_mask", None)
-        current_gradient_accumulation_steps = self.current_gradient_accumulation_steps
+        # Missing when evaluate() runs standalone; eval does not accumulate, so
+        # fall back to 1 to avoid underreporting eval_loss (#2464).
+        current_gradient_accumulation_steps = getattr(
+            self, "current_gradient_accumulation_steps", 1
+        )
         num_processes = self.accelerator.num_processes
 
         input_ids = torch.cat([prompt_ids, completion_ids], dim = 1)
diff --git a/unsloth/models/vision.py b/unsloth/models/vision.py
index 182030c68b..0bd70c6d41 100644
--- a/unsloth/models/vision.py
+++ b/unsloth/models/vision.py
@@ -449,6 +449,23 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
     return output
 
 
+def _missing_torchvision_error(error = None):
+    """True if a VLM processor failed to load due to missing torchvision (#4202).
+
+    Checks availability directly first, then only the specific torchvision-required
+    error text (not any incidental "torchvision" substring like a model path)."""
+    import importlib.util
+
+    if importlib.util.find_spec("torchvision") is None:
+        return True
+    if error is not None:
+        error_str = str(error).lower()
+        return (
+            "requires the torchvision" in error_str or "no module named 'torchvision'" in error_str
+        )
+    return False
+
+
 def _construct_vlm_processor_fallback(tokenizer_name, model_type, token, trust_remote_code):
     """Construct a VLM processor manually when AutoProcessor.from_pretrained fails.
 
@@ -1154,6 +1171,7 @@ class FastBaseModel:
                     except Exception:
                         pass
 
+        _processor_load_error = None
         if (whisper_language and whisper_task) or auto_model.__name__.endswith(
             "ForConditionalGeneration"
         ):
@@ -1166,7 +1184,8 @@ class FastBaseModel:
                     task = whisper_task,
                     trust_remote_code = trust_remote_code,
                 )
-            except Exception:
+            except Exception as e:
+                _processor_load_error = e
                 tokenizer = None
         else:
             try:
@@ -1176,7 +1195,8 @@ class FastBaseModel:
                     token = token,
                     trust_remote_code = trust_remote_code,
                 )
-            except:
+            except Exception as e:
+                _processor_load_error = e
                 tokenizer = get_auto_processor(
                     tokenizer_name,
                     padding_side = "left",
@@ -1200,7 +1220,16 @@ class FastBaseModel:
             )
             if _fallback is not None:
                 tokenizer = _fallback
-            if tokenizer is None:
+            # Missing torchvision silently degrades the VLM processor to a text-only
+            # tokenizer; surface the real cause instead of the later collator error (#4202).
+            if tokenizer is None or not hasattr(tokenizer, "image_processor"):
+                if _missing_torchvision_error(_processor_load_error):
+                    raise ImportError(
+                        f"Unsloth: Could not load the vision processor for `{tokenizer_name}` "
+                        "because torchvision is not installed. transformers requires torchvision "
+                        "for this model's vision (image/video) processors. Please install it, "
+                        "e.g. `pip install torchvision`."
+                    )
                 import sys
                 print(
                     f"Unsloth: Warning - VLM processor fallback returned None for model_type={model_type_arch}",
@@ -1382,6 +1411,7 @@ class FastBaseModel:
         qat_scheme = None,
         target_parameters = None,  # For MoE expert layers (nn.Parameter)
         ensure_weight_tying = False,  # [TODO] Add `ensure_weight_tying` for `modules_to_save` for vision models
+        finetune_audio_layers = False,  # placed last to preserve existing positional argument order
         **kwargs,
     ):
         if os.environ.get("UNSLOTH_ENABLE_FULL_FINETUNING", "0") == "1":
@@ -1400,11 +1430,30 @@ class FastBaseModel:
         if isinstance(model, PeftModelForCausalLM):
             raise RuntimeError("Unsloth: You already added LoRA adapters to your model!")
 
+        # Remember whether the CALLER explicitly opted into audio. "all-linear" turns
+        # the flag on implicitly below, but an old unsloth_zoo that cannot do audio
+        # must not make a plain all-linear (text/vision) run fail.
+        _audio_explicitly_requested = bool(finetune_audio_layers)
         if target_modules == "all-linear":
             finetune_vision_layers = True
             finetune_language_layers = True
             finetune_attention_modules = True
             finetune_mlp_modules = True
+            finetune_audio_layers = True
+        # Older unsloth_zoo (before get_peft_regex gained finetune_audio_layers) does
+        # not accept the kwarg. Pass it only when supported; if the caller EXPLICITLY
+        # asked for audio but it is unsupported, fail loudly rather than silently
+        # training a language-only adapter. (all-linear's implicit opt-in degrades
+        # gracefully instead of raising.)
+        if "finetune_audio_layers" in inspect.signature(get_peft_regex).parameters:
+            _audio_kwargs = {"finetune_audio_layers": finetune_audio_layers}
+        elif _audio_explicitly_requested:
+            raise RuntimeError(
+                "Unsloth: finetune_audio_layers=True requires a newer unsloth_zoo. "
+                "Please upgrade with `pip install --upgrade --no-deps unsloth_zoo`."
+            )
+        else:
+            _audio_kwargs = {}
         if target_modules is None or target_modules == "all-linear":
             target_modules = get_peft_regex(
                 model,
@@ -1412,20 +1461,28 @@ class FastBaseModel:
                 finetune_language_layers = finetune_language_layers,
                 finetune_attention_modules = finetune_attention_modules,
                 finetune_mlp_modules = finetune_mlp_modules,
+                **_audio_kwargs,
             )
         else:
             assert type(target_modules) in (list, tuple, str)
-            if type(target_modules) in (list, tuple) and (
+            # Route an explicit list through get_peft_regex when the caller scoped a
+            # layer family (one of the finetune_* below is off) OR opted into audio (so
+            # the new audio/embedder branches are considered). finetune_audio_layers is
+            # a POSITIVE term here: using `not finetune_audio_layers` would -- since it
+            # defaults False -- force every explicit list through the filter.
+            _scoping = (
                 not finetune_vision_layers
                 or not finetune_language_layers
                 or not finetune_attention_modules
                 or not finetune_mlp_modules
-            ):
-                print(
-                    "Unsloth: Explicit target_modules are constrained by the "
-                    "finetune_(vision|language|attention|mlp) filters; adapters "
-                    "attach only where both select."
-                )
+            )
+            if type(target_modules) in (list, tuple) and (_scoping or finetune_audio_layers):
+                if _scoping:
+                    print(
+                        "Unsloth: Explicit target_modules are constrained by the "
+                        "finetune_(vision|language|attention|mlp) filters; adapters "
+                        "attach only where both select."
+                    )
                 target_modules = get_peft_regex(
                     model,
                     finetune_vision_layers = finetune_vision_layers,
@@ -1433,6 +1490,7 @@ class FastBaseModel:
                     finetune_attention_modules = finetune_attention_modules,
                     finetune_mlp_modules = finetune_mlp_modules,
                     target_modules = list(target_modules),
+                    **_audio_kwargs,
                 )
 
         if hasattr(model, "vllm_engine"):
diff --git a/unsloth/tokenizer_utils.py b/unsloth/tokenizer_utils.py
index f454d5c504..3cec002012 100644
--- a/unsloth/tokenizer_utils.py
+++ b/unsloth/tokenizer_utils.py
@@ -626,6 +626,76 @@ def _load_correct_tokenizer(
         return fast_tokenizer
 
 
+# Qwen3 text models share Qwen3-VL's vocab, so configs ship a vision pad_token;
+# padding text-only training with one yields NaN losses (#3155, #4104).
+_VISION_PAD_TOKENS = frozenset(
+    (
+        "<|vision_pad|>",
+        "<|image_pad|>",
+        "<|video_pad|>",
+        "<|audio_pad|>",
+    )
+)
+# Preference order; <unk> excluded since reusing it as pad masks real OOV tokens.
+_SAFE_TEXT_PAD_TOKENS = ("<|endoftext|>", "<pad>", "[PAD]")
+
+
+def _fix_vision_pad_token(tokenizer):
+    """Swap a vision pad_token on a text-only tokenizer for a safe text token (#3155)."""
+    if tokenizer is None:
+        return tokenizer
+    if hasattr(tokenizer, "image_processor"):
+        return tokenizer
+    pad_token = getattr(tokenizer, "pad_token", None)
+    if pad_token is None or pad_token not in _VISION_PAD_TOKENS:
+        return tokenizer
+
+    get_vocab = getattr(tokenizer, "get_vocab", None)
+    if get_vocab is None:
+        return tokenizer
+    vocab = get_vocab()
+    if not isinstance(vocab, dict):
+        return tokenizer
+    new_pad_token = None
+    for candidate in _SAFE_TEXT_PAD_TOKENS:
+        if candidate in vocab and candidate != getattr(tokenizer, "eos_token", None):
+            new_pad_token = candidate
+            break
+    # Fall back to eos_token only if it is a distinct, non-vision token.
+    if new_pad_token is None:
+        eos_token = getattr(tokenizer, "eos_token", None)
+        if eos_token is not None and eos_token != pad_token and eos_token not in _VISION_PAD_TOKENS:
+            new_pad_token = eos_token
+    if new_pad_token is None:
+        return tokenizer
+
+    tokenizer.pad_token = new_pad_token
+    logger.warning(
+        f"Unsloth: pad_token was a vision token ({pad_token}) on a text-only "
+        f"model. Replaced with {new_pad_token} to avoid NaN losses."
+    )
+    return tokenizer
+
+
+def _fix_pad_token(tokenizer):
+    """Heal a bad/missing pad_token before chat-template repair.
+
+    Delegates to unsloth_zoo's shared fix_pad_token (single source of truth) when
+    available, falling back to the narrow vision-token swap against an older
+    unsloth_zoo. allow_add=False keeps this side-effect free: there is no model
+    here to resize embeddings, so a brand new pad token is never added - the later
+    model-aware patch_tokenizer call finishes the job and is idempotent.
+    """
+    if tokenizer is None:
+        return tokenizer
+    try:
+        from unsloth_zoo.pad_token import fix_pad_token
+    except Exception:
+        return _fix_vision_pad_token(tokenizer)
+    fix_pad_token(tokenizer, allow_add = False)
+    return tokenizer
+
+
 def load_correct_tokenizer(
     tokenizer_name,
     model_max_length = None,
@@ -645,6 +715,9 @@ def load_correct_tokenizer(
         fix_tokenizer = fix_tokenizer,
     )
 
+    if fix_tokenizer:
+        _fix_pad_token(tokenizer)
+
     ### 1. Fixup tokenizer's chat_template
     old_chat_template = getattr(tokenizer, "chat_template", None)
 
diff --git a/unsloth/utils/attention_dispatch.py b/unsloth/utils/attention_dispatch.py
index 3ff581b7fc..2e984bad0a 100644
--- a/unsloth/utils/attention_dispatch.py
+++ b/unsloth/utils/attention_dispatch.py
@@ -137,6 +137,25 @@ def run_attention(
     requires_grad = context.requires_grad
     sliding_window = context.sliding_window
 
+    # DoRA promotes q/k/v_proj outputs to fp32, which FlashAttention rejects, so
+    # downcast any fp32 Q/K/V to a flash-supported dtype (#1013).
+    if backend in (FLASH_DENSE, FLASH_VARLEN) and torch.float32 in (
+        Q.dtype,
+        K.dtype,
+        V.dtype,
+    ):
+        # Prefer the autocast dtype, else a non-fp32 input's dtype, then clamp.
+        if torch.is_autocast_enabled():
+            try:
+                flash_dtype = torch.get_autocast_dtype("cuda")
+            except (AttributeError, TypeError):
+                flash_dtype = torch.get_autocast_gpu_dtype()
+        else:
+            flash_dtype = next((d for d in (Q.dtype, K.dtype, V.dtype) if d != torch.float32), None)
+        if flash_dtype not in (torch.float16, torch.bfloat16):
+            flash_dtype = torch.bfloat16 if SUPPORTS_BFLOAT16 else torch.float16
+        Q, K, V = Q.to(flash_dtype), K.to(flash_dtype), V.to(flash_dtype)
+
     if backend == FLASH_VARLEN:
         Q_f = Q.transpose(1, 2).reshape(bsz * q_len, n_heads, head_dim)
         K_f = K.transpose(1, 2).reshape(bsz * q_len, config.n_kv_heads, head_dim)
diff --git a/unsloth_cli/commands/studio.py b/unsloth_cli/commands/studio.py
index cde3d2cdbc..5975e7e355 100644
--- a/unsloth_cli/commands/studio.py
+++ b/unsloth_cli/commands/studio.py
@@ -776,6 +776,16 @@ def studio_default(
                 err = True,
             )
             raise typer.Exit(2)
+        # Same for --api-only: dropping it here would silently serve the UI.
+        if api_only:
+            typer.echo(
+                f"Error: --api-only on `unsloth studio` applies to the "
+                f"plain-server path only. For `unsloth studio "
+                f"{ctx.invoked_subcommand}`, put it after the subcommand: "
+                f"`unsloth studio {ctx.invoked_subcommand} --api-only ...`",
+                err = True,
+            )
+            raise typer.Exit(2)
         return
 
     # --secure requires the tunnel; force a loopback bind.
@@ -1031,6 +1041,12 @@ def run(
     host: str = typer.Option("127.0.0.1", "--host", "-H"),
     # `-f` removed (clustered `-fa`/`-fit*`); studio_default keeps it.
     frontend: Optional[Path] = typer.Option(None, "--frontend"),
+    api_only: bool = typer.Option(
+        False,
+        "--api-only",
+        help = "Serve only the API (no web UI), for a headless model server. "
+        "Pairs with --secure to expose the API over the Cloudflare link alone.",
+    ),
     silent: bool = typer.Option(False, "--silent", "-q"),
     enable_tools: Optional[bool] = typer.Option(
         None,
@@ -1214,6 +1230,8 @@ def run(
         args.append("--load-in-4bit" if load_in_4bit else "--no-load-in-4bit")
         if frontend:
             args.extend(["--frontend", str(frontend)])
+        if api_only:
+            args.append("--api-only")
         if silent:
             args.append("--silent")
         # Forward the resolved tool policy so the child doesn't re-resolve.
@@ -1262,9 +1280,13 @@ def run(
         host = host,
         port = port,
         silent = True,
+        api_only = api_only,
         llama_parallel_slots = parallel,
         cloudflare = cloudflare,
         secure = secure,
+        # Headless serving prints its own URL/API-key banner; the Tauri-only
+        # TAURI_PORT line would corrupt that machine-parseable output.
+        emit_tauri_port = False,
     )
     if frontend is not None:
         run_kwargs["frontend_path"] = frontend
diff --git a/unsloth_cli/tests/test_studio_run_parallel_flag.py b/unsloth_cli/tests/test_studio_run_parallel_flag.py
index 8caf31a432..558b268a4d 100644
--- a/unsloth_cli/tests/test_studio_run_parallel_flag.py
+++ b/unsloth_cli/tests/test_studio_run_parallel_flag.py
@@ -361,6 +361,29 @@ def test_studio_default_rejects_parallel_when_subcommand_invoked():
     ), f"error message must show the corrected invocation; got: {combined}"
 
 
+def test_studio_default_rejects_api_only_when_subcommand_invoked():
+    """`unsloth studio --api-only run ...` would silently serve the UI (the
+    parent's --api-only never reaches run). The callback rejects with exit 2
+    and points at the subcommand flag."""
+    studio_mod = _load_run_command()
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.add_typer(studio_mod.studio_app, name = "studio")
+
+    runner = CliRunner()
+    result = runner.invoke(app, ["studio", "--api-only", "run", "--model", "X"])
+    assert result.exit_code == 2, (
+        f"expected exit 2 when --api-only is on studio group with a "
+        f"subcommand invoked; got {result.exit_code}; output={result.output!r}"
+    )
+    combined = (result.output or "") + (getattr(result, "stderr", "") or "")
+    assert "--api-only" in combined, combined
+    assert (
+        "run --api-only" in combined
+    ), f"error message must show the corrected invocation; got: {combined}"
+
+
 def test_studio_default_default_parallel_with_subcommand_does_not_error():
     """Omitting --parallel on the group must still let subcommands
     run; the group's default 1 is benign."""
@@ -445,3 +468,81 @@ def test_in_venv_path_passes_parallel_to_run_server(monkeypatch, value):
     assert (
         captured.get("llama_parallel_slots") == value
     ), f"run_server got llama_parallel_slots={captured.get('llama_parallel_slots')!r}, expected {value}"
+
+
+# --api-only: serve API only (no UI). Both re-exec and in-venv paths must carry it.
+
+
+def test_api_only_option_is_registered():
+    studio_mod = _load_run_command()
+    import inspect
+
+    opt = inspect.signature(studio_mod.run).parameters["api_only"].default
+    assert "--api-only" in set(getattr(opt, "param_decls", []) or [])
+    assert getattr(opt, "default", None) is False  # opt-in; plain run keeps the UI
+
+
+@pytest.mark.parametrize(
+    "extra,present",
+    [
+        (["--api-only"], True),
+        (["--secure", "--api-only"], True),  # secure headless path
+        ([], False),
+    ],
+)
+def test_reexec_forwards_api_only(monkeypatch, extra, present):
+    """`--api-only` (and only when typed) must reach the re-exec'd child."""
+    result, captured = _invoke_run(monkeypatch, _BASE + extra)
+    assert len(captured) == 1, result.output
+    argv = captured[0]["argv"]
+    assert ("--api-only" in argv) is present, argv
+
+
+@pytest.mark.parametrize("extra,expected", [(["--api-only"], True), ([], False)])
+def test_in_venv_path_passes_api_only_to_run_server(monkeypatch, extra, expected):
+    """In-venv path must forward --api-only to run_server(api_only=...)."""
+    studio_mod = _load_run_command()
+
+    fake_venv = Path("/fake/studio/venv/unsloth_studio")
+    monkeypatch.setattr(sys, "prefix", str(fake_venv))
+    monkeypatch.setattr(studio_mod, "STUDIO_HOME", fake_venv.parent)
+
+    from unsloth_cli import _tool_policy as _tp_mod
+
+    monkeypatch.setattr(
+        _tp_mod,
+        "resolve_tool_policy",
+        lambda host, flag, yes, silent: False if flag is None else bool(flag),
+    )
+
+    captured: dict = {}
+
+    def fake_run_server(**kwargs):
+        captured.update(kwargs)
+        raise _RunServerCaptured(kwargs)
+
+    fake_backend_run = sys.modules.setdefault(
+        "studio.backend.run", _types_module("studio.backend.run")
+    )
+    fake_backend_run.run_server = fake_run_server
+    fake_backend_run._resolve_external_ip = lambda: "127.0.0.1"
+    monkeypatch.setattr(studio_mod, "_RUN_MODULE", fake_backend_run)
+
+    import typer as _typer
+
+    app = _typer.Typer()
+    app.command(
+        context_settings = {
+            "allow_extra_args": True,
+            "ignore_unknown_options": True,
+        },
+    )(studio_mod.run)
+    CliRunner().invoke(app, _BASE + extra, catch_exceptions = True)
+
+    assert (
+        captured.get("api_only") is expected
+    ), f"run_server got api_only={captured.get('api_only')!r}, expected {expected}"
+    # Headless serving must suppress the Tauri-only TAURI_PORT line.
+    assert (
+        captured.get("emit_tauri_port") is False
+    ), f"run_server got emit_tauri_port={captured.get('emit_tauri_port')!r}, expected False"