Merge remote-tracking branch 'origin/main' into dg-onto-main

install.ps1

@@ -1876,7 +1876,7 @@ shell.Run cmd, 0, False
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.2" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.3" unsloth-zoo }
             if ($baseInstallExit -eq 0) {
                 # Resolve pydantic WITH deps so pip pins pydantic-core
                 # to the matching version (no-torch-runtime.txt below
@@ -1890,7 +1890,7 @@ shell.Run cmd, 0, False
                 }
             }
         } else {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.2" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --reinstall-package unsloth --reinstall-package unsloth-zoo "unsloth>=2026.6.3" unsloth-zoo }
         }
         if ($baseInstallExit -ne 0) {
             Write-Host "[ERROR] Failed to install unsloth (exit code $baseInstallExit)" -ForegroundColor Red
@@ -1937,7 +1937,7 @@ shell.Run cmd, 0, False
         if ($SkipTorch) {
             # No-torch: install unsloth + unsloth-zoo with --no-deps, then
             # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.2" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --no-deps --upgrade-package unsloth --upgrade-package unsloth-zoo "unsloth>=2026.6.3" unsloth-zoo }
             if ($baseInstallExit -eq 0) {
                 # Same pydantic-with-deps trick as the migrated branch.
                 $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython pydantic }
@@ -1949,7 +1949,7 @@ shell.Run cmd, 0, False
                 }
             }
         } elseif ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.2" unsloth-zoo }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth "unsloth>=2026.6.3" unsloth-zoo }
         } else {
             $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython --upgrade-package unsloth -- "$PackageName" }
         }
@@ -1977,7 +1977,7 @@ shell.Run cmd, 0, False
         Write-TauriLog "STEP" "Installing unsloth"
         substep "installing unsloth (this may take a few minutes)..."
         if ($StudioLocalInstall) {
-            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython unsloth-zoo "unsloth>=2026.6.2" --torch-backend=auto }
+            $baseInstallExit = Invoke-InstallCommand { uv pip install --python $VenvPython unsloth-zoo "unsloth>=2026.6.3" --torch-backend=auto }
             if ($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)

install.sh

@@ -2405,7 +2405,7 @@ if [ "$_MIGRATED" = true ]; then
         # to prevent transitive torch resolution.
         run_install_cmd "install unsloth (migrated no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.2" unsloth-zoo
+            "unsloth>=2026.6.3" unsloth-zoo
         # 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.
@@ -2418,7 +2418,7 @@ if [ "$_MIGRATED" = true ]; then
     else
         run_install_cmd "install unsloth (migrated)" uv pip install --python "$_VENV_PY" \
             --reinstall-package unsloth --reinstall-package unsloth-zoo \
-            "unsloth>=2026.6.2" unsloth-zoo
+            "unsloth>=2026.6.3" unsloth-zoo
     fi
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         substep "overlaying local repo (editable)..."
@@ -2622,7 +2622,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         # runtime deps (typer, safetensors, transformers, etc.) with --no-deps.
         run_install_cmd "install unsloth (no-torch)" uv pip install --python "$_VENV_PY" --no-deps \
             --upgrade-package unsloth --upgrade-package unsloth-zoo \
-            "unsloth>=2026.6.2" unsloth-zoo
+            "unsloth>=2026.6.3" unsloth-zoo
         # Same pydantic-with-deps trick as the migrated branch.
         run_install_cmd "install pydantic (with deps for compatible core)" \
             uv pip install --python "$_VENV_PY" pydantic
@@ -2640,7 +2640,7 @@ elif [ -n "$TORCH_INDEX_URL" ]; then
         fi
     elif [ "$STUDIO_LOCAL_INSTALL" = true ]; then
         run_install_cmd "install unsloth (local)" uv pip install --python "$_VENV_PY" \
-            --upgrade-package unsloth "unsloth>=2026.6.2" unsloth-zoo
+            --upgrade-package unsloth "unsloth>=2026.6.3" unsloth-zoo
         substep "overlaying local repo (editable)..."
         run_install_cmd "overlay local repo" uv pip install --python "$_VENV_PY" -e "$_REPO_ROOT" --no-deps
         substep "overlaying unsloth-zoo from git main..."
@@ -2672,7 +2672,7 @@ else
     tauri_log "STEP" "Installing Unsloth"
     substep "installing unsloth (this may take a few minutes)..."
     if [ "$STUDIO_LOCAL_INSTALL" = true ]; then
-        run_install_cmd "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" unsloth-zoo "unsloth>=2026.6.2" --torch-backend=auto
+        run_install_cmd "install unsloth (auto torch backend)" uv pip install --python "$_VENV_PY" unsloth-zoo "unsloth>=2026.6.3" --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..."

pyproject.toml

@@ -70,7 +70,7 @@ triton = [
 ]
 
 huggingfacenotorch = [
-    "unsloth_zoo>=2026.6.2",
+    "unsloth_zoo>=2026.6.3",
     "wheel>=0.42.0",
     "packaging",
     "numpy",
@@ -91,7 +91,7 @@ huggingfacenotorch = [
 ]
 huggingface = [
     "unsloth[huggingfacenotorch]",
-    "unsloth_zoo>=2026.6.2",
+    "unsloth_zoo>=2026.6.3",
     "torchvision",
     "unsloth[triton]",
 ]
@@ -581,7 +581,7 @@ colab-ampere-torch220 = [
     "flash-attn>=2.6.3 ; ('linux' in sys_platform)",
 ]
 colab-new = [
-    "unsloth_zoo>=2026.6.2",
+    "unsloth_zoo>=2026.6.3",
     "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",

studio/backend/core/inference/inference.py

@@ -25,6 +25,7 @@ from utils.hardware import (
     get_visible_gpu_count,
 )
 from core.inference.audio_codecs import AudioCodecManager
+from core.inference.runtime_context import runtime_context_length
 from io import StringIO
 import structlog
 from loggers import get_logger
@@ -405,6 +406,10 @@ class InferenceBackend:
 
                 # Reject CPU/disk offload for audio models too
                 raise_if_offloaded(self.models[model_name]["model"], device_map, "Inference")
+                self.models[model_name]["context_length"] = runtime_context_length(
+                    self.models[model_name].get("model"),
+                    max_seq_length,
+                )
 
                 self.active_model_name = model_name
                 self.loading_models.discard(model_name)
@@ -485,6 +490,10 @@ class InferenceBackend:
                 self.models[model_name]["tokenizer"] = tokenizer
 
             raise_if_offloaded(self.models[model_name]["model"], device_map, "Inference")
+            self.models[model_name]["context_length"] = runtime_context_length(
+                self.models[model_name].get("model"),
+                max_seq_length,
+            )
 
             self._load_chat_template_info(model_name)
 

studio/backend/core/inference/mlx_inference.py

@@ -7,6 +7,7 @@ instead of torch/transformers for model loading and generation.
 
 import threading
 from typing import Optional, Generator
+from core.inference.runtime_context import runtime_context_length
 from loggers import get_logger
 
 logger = get_logger(__name__)
@@ -175,6 +176,7 @@ class MLXInferenceBackend:
             "is_audio": False,
             "audio_type": None,
             "has_audio_input": False,
+            "context_length": runtime_context_length(self._model, max_seq_length),
         }
         # Capture chat_template_info so the worker IPC reply ships it back and
         # the route layer classifies capabilities like the other paths.

studio/backend/core/inference/orchestrator.py

@@ -727,6 +727,7 @@ class InferenceOrchestrator:
                         "is_audio": model_info.get("is_audio", False),
                         "audio_type": model_info.get("audio_type"),
                         "has_audio_input": model_info.get("has_audio_input", False),
+                        "context_length": model_info.get("context_length"),
                     }
                     # Mirror chat_template_info so routes can classify caps
                     # without re-entering the subprocess.

studio/backend/core/inference/runtime_context.py

@@ -0,0 +1,22 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
+
+"""Runtime context length helpers shared by inference backends."""
+
+from __future__ import annotations
+
+from typing import Any, Optional
+
+
+def runtime_context_length(model: Any, fallback: Optional[int] = None) -> Optional[int]:
+    """Return the effective context length Unsloth attached to a loaded model."""
+    for value in (getattr(model, "max_seq_length", None), fallback):
+        if isinstance(value, bool):
+            continue
+        try:
+            value_int = int(value)
+        except (TypeError, ValueError):
+            continue
+        if value_int > 0:
+            return value_int
+    return None

studio/backend/core/inference/worker.py

@@ -315,6 +315,18 @@ def _handle_load(backend, config: dict, resp_queue: Any) -> None:
                 "audio_type": getattr(mc, "audio_type", None),
                 "has_audio_input": getattr(mc, "has_audio_input", False),
             }
+            try:
+                _bm = getattr(backend, "models", {}) or {}
+                _entry = (
+                    _bm.get(mc.identifier)
+                    or _bm.get(getattr(backend, "active_model_name", None))
+                    or {}
+                )
+                _context_length = _entry.get("context_length")
+                if _context_length is not None:
+                    model_info["context_length"] = int(_context_length)
+            except Exception as _ctx_exc:
+                logger.warning("context_length forward failed: %s", _ctx_exc)
             # Forward chat_template_info so the parent can classify capabilities.
             try:
                 _bm = getattr(backend, "models", {}) or {}
@@ -881,6 +893,7 @@ def run_inference_process(*, cmd_queue: Any, resp_queue: Any, cancel_event, conf
                             name: {
                                 "is_vision": info.get("is_vision", False),
                                 "is_lora": info.get("is_lora", False),
+                                "context_length": info.get("context_length"),
                             }
                             for name, info in backend.models.items()
                         },

studio/backend/models/inference.py

@@ -174,7 +174,7 @@ class LoadResponse(BaseModel):
         description = "Whether the model defaults require trust_remote_code to be enabled for loading.",
     )
     context_length: Optional[int] = Field(
-        None, description = "Model's native context length (from GGUF metadata)"
+        None, description = "Runtime context length in tokens for the loaded model"
     )
     max_context_length: Optional[int] = Field(
         None, description = "Maximum context length currently available on this hardware"

studio/backend/routes/inference.py

@@ -29,6 +29,16 @@ from utils.models import extract_model_size_b as _extract_model_size_b
 from utils.api_errors import openai_error_body, anthropic_error_body
 
 
+def _positive_int_or_none(value: Any) -> Optional[int]:
+    if isinstance(value, bool):
+        return None
+    try:
+        value_int = int(value)
+    except (TypeError, ValueError):
+        return None
+    return value_int if value_int > 0 else None
+
+
 def _install_httpcore_asyncgen_silencer() -> None:
     """Silence benign httpx/httpcore asyncgen GC noise on Python 3.13.
 
@@ -1391,6 +1401,7 @@ async def load_model(
                     reasoning_always_on = _sf_flags["reasoning_always_on"],
                     supports_preserve_thinking = _sf_flags["supports_preserve_thinking"],
                     supports_tools = _sf_flags["supports_tools"],
+                    context_length = _positive_int_or_none(_model_info.get("context_length")),
                     chat_template = _chat_template,
                 )
 
@@ -1733,6 +1744,7 @@ async def load_model(
             reasoning_always_on = _sf_flags["reasoning_always_on"],
             supports_preserve_thinking = _sf_flags["supports_preserve_thinking"],
             supports_tools = _sf_flags["supports_tools"],
+            context_length = _positive_int_or_none(_model_info.get("context_length")),
             chat_template = _chat_template,
         )
 
@@ -2122,6 +2134,7 @@ async def get_status(current_subject: str = Depends(get_current_subject)):
             reasoning_always_on = _sf_flags["reasoning_always_on"],
             supports_preserve_thinking = _sf_flags["supports_preserve_thinking"],
             supports_tools = _sf_flags["supports_tools"],
+            context_length = _positive_int_or_none(model_info.get("context_length")),
             chat_template = chat_template,
             llama_cpp_supports_mtp = _supports_mtp,
             llama_cpp_prebuilt_stale = _stale,
@@ -4628,28 +4641,38 @@ def _openai_model_objects() -> list[dict]:
             "created": _created,
             "owned_by": "local",
         }
-        # Extension fields: the real per-request window (post /props readback)
-        # so clients can budget/compact against the enforced limit.
-        if llama_backend.context_length:
-            entry["context_length"] = llama_backend.context_length
-        if llama_backend.max_context_length:
-            entry["max_context_length"] = llama_backend.max_context_length
+        _ctx = _positive_int_or_none(getattr(llama_backend, "context_length", None))
+        if _ctx is not None:
+            entry["context_length"] = _ctx
+        _max_ctx = _positive_int_or_none(getattr(llama_backend, "max_context_length", None))
+        if _max_ctx is not None:
+            entry["max_context_length"] = _max_ctx
+        _native_ctx = _positive_int_or_none(getattr(llama_backend, "native_context_length", None))
+        if _native_ctx is not None:
+            entry["native_context_length"] = _native_ctx
         models.append(entry)
 
     # Check Unsloth backend
     backend = get_inference_backend()
     if backend.active_model_name:
+        model_info = backend.models.get(backend.active_model_name, {})
         entry = {
             "id": backend.active_model_name,
             "object": "model",
             "created": _created,
             "owned_by": "local",
         }
-        _sf_ctx = getattr(backend, "context_length", None) or getattr(
-            backend, "max_seq_length", None
-        )
-        if _sf_ctx:
-            entry["context_length"] = _sf_ctx
+        _ctx = _positive_int_or_none(model_info.get("context_length"))
+        if _ctx is None:
+            for _candidate in (
+                getattr(backend, "context_length", None),
+                getattr(backend, "max_seq_length", None),
+            ):
+                _ctx = _positive_int_or_none(_candidate)
+                if _ctx is not None:
+                    break
+        if _ctx is not None:
+            entry["context_length"] = _ctx
         models.append(entry)
 
     return models

studio/backend/routes/llama.py

@@ -31,6 +31,7 @@ class LlamaUpdateJob(BaseModel):
     from_tag: Optional[str] = None
     to_tag: Optional[str] = None
     error: Optional[str] = None
+    progress: Optional[float] = Field(None, description = "0..1 while running, 1 on success.")
     started_at: Optional[str] = None
     finished_at: Optional[str] = None
 
@@ -40,7 +41,9 @@ class LlamaUpdateStatusResponse(BaseModel):
         False,
         description = "True when the install came from an Unsloth prebuilt (has a marker).",
     )
-    update_available: bool = Field(False, description = "True when installed_tag != latest_tag.")
+    update_available: bool = Field(
+        False, description = "True when the latest release is genuinely newer than the install."
+    )
     stale: bool = Field(
         False, description = "Update available AND install older than the staleness threshold."
     )

studio/backend/run.py

@@ -232,6 +232,9 @@ def _verify_global_reachability(display_host: str, port: int) -> None:
     public internet. Synchronous so output lands between the banner URLs and the
     stop hint. Bounded at ~15s; failures swallowed (verifier failing != Studio
     failing). Only meaningful for a wildcard bind."""
+    global _public_reachable
+    # Reset to "unknown" each run; set True/False only when the probe decides.
+    _public_reachable = None
     import ipaddress
     import json
     import time
@@ -324,12 +327,14 @@ def _verify_global_reachability(display_host: str, port: int) -> None:
 
         print("", flush = True)
         if ok_nodes:
+            _public_reachable = True
             print(
                 f"{ok_c}  Reachability check: {url}/ is reachable from the "
                 f"public internet ({ok_nodes}/{total} probe nodes connected).{reset}",
                 flush = True,
             )
         elif err_nodes:
+            _public_reachable = False
             print(
                 f"{err_c}  Reachability check: {url}/ is NOT reachable from "
                 f"the public internet ({err_nodes}/{total} probe nodes failed).{reset}",
@@ -422,7 +427,9 @@ def _print_cloudflare_line() -> None:
     """Print the Cloudflare quick-tunnel URL for 0.0.0.0 binds, if one is up.
 
     Reads the module-level URL set by ``run_server``. Prints nothing when the
-    tunnel is disabled or failed -- failures are silently ignored.
+    tunnel is disabled or failed -- failures are silently ignored. When the public
+    reachability probe just failed (``_public_reachable is False``) but the tunnel
+    is up, reword to point the user at the Cloudflare link as the way in.
     """
     if not _cloudflare_url:
         return
@@ -430,7 +437,10 @@ def _print_cloudflare_line() -> None:
 
     accent = "\033[38;5;150;1m"
     reset = "\033[0m"
-    line = f"  Secure link access via Cloudflare: {_cloudflare_url}"
+    if _public_reachable is False:
+        line = f"  Use the secure link access via Cloudflare instead: {_cloudflare_url}"
+    else:
+        line = f"  Secure link access via Cloudflare: {_cloudflare_url}"
     print(f"{accent}{line}{reset}" if stdout_supports_color() else line)
 
 
@@ -622,6 +632,12 @@ _shutdown_event = None
 # None when there is no tunnel (loopback, disabled, or a silently-ignored failure).
 _cloudflare_url = None
 
+# Public reachability from the last _verify_global_reachability run, read by the
+# Cloudflare banner line. True when the public ip:port probe confirmed reachable,
+# False when it confirmed NOT reachable, None when the probe did not run or could
+# not decide (timeout, blocked, private address).
+_public_reachable = None
+
 
 _DEFAULT_FRONTEND_PATH = Path(__file__).resolve().parent.parent / "frontend" / "dist"
 

studio/backend/tests/test_cloudflare_tunnel.py

@@ -13,6 +13,7 @@ import importlib.util
 import io
 import sys
 import tarfile
+import types
 from pathlib import Path
 
 import pytest
@@ -423,3 +424,56 @@ def test_run_server_gates_tunnel_on_wildcard():
     source = _RUN_PY.read_text()
     assert "_cloudflare_enabled" in source
     assert 'host == "0.0.0.0"' in source
+
+
+def _run_print_cloudflare_line(monkeypatch, *, cloudflare_url, public_reachable):
+    """Exec the real _print_cloudflare_line source in isolation (run.py has heavy
+    deps), with the two module globals injected and startup_banner stubbed."""
+    src = _RUN_PY.read_text()
+    tree = ast.parse(src)
+    func_src = next(
+        ast.get_source_segment(src, n)
+        for n in ast.walk(tree)
+        if isinstance(n, ast.FunctionDef) and n.name == "_print_cloudflare_line"
+    )
+    stub = types.ModuleType("startup_banner")
+    stub.stdout_supports_color = lambda: False
+    monkeypatch.setitem(sys.modules, "startup_banner", stub)
+    captured: list[str] = []
+    ns = {
+        "_cloudflare_url": cloudflare_url,
+        "_public_reachable": public_reachable,
+        "print": lambda *a, **k: captured.append(" ".join(str(x) for x in a)),
+    }
+    exec(compile(func_src, "<print_cloudflare_line>", "exec"), ns)
+    ns["_print_cloudflare_line"]()
+    return "\n".join(captured)
+
+
+def test_cloudflare_line_reworded_when_public_unreachable(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch, cloudflare_url = "https://x.trycloudflare.com", public_reachable = False
+    )
+    assert "Use the secure link access via Cloudflare instead: https://x.trycloudflare.com" in out
+
+
+def test_cloudflare_line_default_wording_when_reachable(monkeypatch):
+    out = _run_print_cloudflare_line(
+        monkeypatch, cloudflare_url = "https://x.trycloudflare.com", public_reachable = True
+    )
+    assert "Secure link access via Cloudflare: https://x.trycloudflare.com" in out
+    assert "Use the secure link" not in out
+
+
+def test_cloudflare_line_default_wording_when_unknown(monkeypatch):
+    # Probe did not run / could not decide -> keep the existing wording.
+    out = _run_print_cloudflare_line(
+        monkeypatch, cloudflare_url = "https://x.trycloudflare.com", public_reachable = None
+    )
+    assert "Secure link access via Cloudflare: https://x.trycloudflare.com" in out
+    assert "Use the secure link" not in out
+
+
+def test_cloudflare_line_prints_nothing_without_tunnel(monkeypatch):
+    out = _run_print_cloudflare_line(monkeypatch, cloudflare_url = None, public_reachable = False)
+    assert out == ""

studio/backend/tests/test_lemonade_llamacpp_rocm_bins_mock.py

@@ -32,14 +32,19 @@ if resolve_lemonade_rocm_choice is None or _LEMONADE_GFX_FAMILIES is None:
 
 @pytest.fixture(autouse = True)
 def _clear_lemonade_release_cache():
-    """Prevent cross-test pollution of the lemonade release lru_cache when
-    future tests vary the fetch_json mock return value."""
+    """Prevent cross-test pollution of the lemonade release lru_cache and
+    selection-log dedup set when tests vary the fetch_json mock return value."""
     _cache = getattr(_mod, "_fetch_lemonade_release_cached", None)
+    _logged: set | None = getattr(_mod, "_lemonade_selection_logged", None)
     if _cache is not None and hasattr(_cache, "cache_clear"):
         _cache.cache_clear()
+    if _logged is not None:
+        _logged.clear()
     yield
     if _cache is not None and hasattr(_cache, "cache_clear"):
         _cache.cache_clear()
+    if _logged is not None:
+        _logged.clear()
 
 
 _STUB_TAG = "b1262"

studio/backend/tests/test_llama_cpp_freshness.py

@@ -21,12 +21,25 @@ _BACKEND_DIR = str(Path(__file__).resolve().parent.parent)
 if _BACKEND_DIR not in sys.path:
     sys.path.insert(0, _BACKEND_DIR)
 
+
+class _NoopLogger:
+    """structlog-style logger: every method swallows positional + kwargs.
+
+    A stdlib logging.Logger rejects structlog's keyword fields (e.g.
+    ``logger.warning(msg, error=...)``), which leaked into the update module's
+    error path and failed only when this file's stub loaded first.
+    """
+
+    def __getattr__(self, _name):
+        return lambda *a, **k: None
+
+
 _loggers_stub = _types.ModuleType("loggers")
-_loggers_stub.get_logger = lambda name: __import__("logging").getLogger(name)
+_loggers_stub.get_logger = lambda *a, **k: _NoopLogger()
 sys.modules.setdefault("loggers", _loggers_stub)
 
 _structlog_stub = _types.ModuleType("structlog")
-_structlog_stub.get_logger = lambda *a, **k: __import__("logging").getLogger("stub")
+_structlog_stub.get_logger = lambda *a, **k: _NoopLogger()
 sys.modules.setdefault("structlog", _structlog_stub)
 
 import pytest
@@ -51,6 +64,11 @@ def _write_marker(install_dir: Path, **overrides) -> Path:
         .replace("+00:00", "Z"),
     }
     payload.update(overrides)
+    # The installer always writes `tag` and `release_tag` from the same release
+    # (a normalized base vs the full release tag), so keep the pair consistent
+    # when a test overrides only `tag`.
+    if "tag" in overrides and "release_tag" not in overrides:
+        payload["release_tag"] = overrides["tag"]
     install_dir.mkdir(parents = True, exist_ok = True)
     (install_dir / "UNSLOTH_PREBUILT_INFO.json").write_text(json.dumps(payload))
     return install_dir / "UNSLOTH_PREBUILT_INFO.json"
@@ -303,3 +321,115 @@ def test_format_stale_warning_singular_day():
     msg = fr.format_stale_warning({"installed_tag": "b9190", "latest_tag": "b9300", "age_days": 1})
     assert "1 day" in msg
     assert "1 days" not in msg
+
+
+# parse_base_build / is_behind.
+
+
+def test_parse_base_build():
+    assert fr.parse_base_build("b9596") == 9596
+    assert fr.parse_base_build(" b9596 ") == 9596
+    assert fr.parse_base_build("b9596-mix-e6f2453") == 9596  # mix suffix doesn't defeat it
+    assert fr.parse_base_build("9596") is None
+    assert fr.parse_base_build("master-abc") is None
+    assert fr.parse_base_build("") is None
+    assert fr.parse_base_build(None) is None
+
+
+@pytest.mark.parametrize(
+    "installed, latest, expected",
+    [
+        (
+            "b9596-mix-e6f2453",
+            "b9596-mix-e6f2453",
+            False,
+        ),  # already on the mix latest -> not behind
+        ("b9596", "b9594", False),  # latest is an older build -> downgrade guard
+        ("b9596", "b9594-mix-xxx", False),  # older mix latest -> still guarded
+        ("b9500", "b9596-mix-e6f2453", True),  # newer base -> behind
+        ("b9596-mix-aaa", "b9596-mix-bbb", True),  # new mix at same base -> behind
+        ("b9596", "b9596-mix-bbb", True),  # clean -> mix at same base -> behind
+        ("b9596-mix-aaa", "b9596", False),  # bare base never supersedes a mix install
+        ("b9596", "b9596", False),  # identical -> not behind
+        (" b9596 ", "b9596", False),  # whitespace-only diff -> not behind
+        ("master-abc", "master-def", True),  # non-bNNNN both -> plain inequality
+        ("master-abc", "master-abc", False),
+        (None, "b9596", False),
+        ("b9596", None, False),
+    ],
+)
+def test_is_behind(installed, latest, expected):
+    assert fr.is_behind(installed, latest) is expected
+
+
+def test_check_prebuilt_freshness_not_behind_on_mix_latest(monkeypatch, tmp_path):
+    # Installed the mix latest: marker base tag b9596, full release_tag with sha,
+    # GitHub latest is that same full tag. Must not report behind (sticky bug).
+    install_dir = tmp_path / "llama.cpp"
+    _write_marker(install_dir, tag = "b9596", release_tag = "b9596-mix-e6f2453")
+    bin_path = _fake_binary(install_dir, layout = "root")
+    monkeypatch.setattr(
+        fr, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9596-mix-e6f2453"
+    )
+    info = fr.check_prebuilt_freshness(str(bin_path))
+    assert info["behind"] is False
+    assert info["stale"] is False
+
+
+def test_check_prebuilt_freshness_downgrade_guard(monkeypatch, tmp_path):
+    # A lagging latest (older build than installed) must never read as behind/stale.
+    install_dir = tmp_path / "llama.cpp"
+    _write_marker(
+        install_dir,
+        tag = "b9585",
+        installed_at_utc = (datetime.now(tz = timezone.utc) - timedelta(days = 30))
+        .isoformat()
+        .replace("+00:00", "Z"),
+    )
+    bin_path = _fake_binary(install_dir, layout = "root")
+    monkeypatch.setattr(fr, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+    info = fr.check_prebuilt_freshness(str(bin_path))
+    assert info["behind"] is False
+    assert info["stale"] is False
+
+
+def test_fetch_latest_release_tag_uses_publish_time(monkeypatch):
+    # Resolves newest by published_at (like the installer), skips drafts/prereleases,
+    # and does NOT just take GitHub's first/`/releases/latest` item.
+    import urllib.request
+
+    class _Resp:
+        def __init__(self, payload):
+            self._p = json.dumps(payload).encode()
+
+        def read(self):
+            return self._p
+
+        def __enter__(self):
+            return self
+
+        def __exit__(self, *a):
+            return False
+
+    payload = [
+        {
+            "tag_name": "b9518",
+            "draft": False,
+            "prerelease": False,
+            "published_at": "2026-06-04T21:11:19Z",
+        },
+        {
+            "tag_name": "b9596-mix-e6f2453",
+            "draft": False,
+            "prerelease": False,
+            "published_at": "2026-06-11T22:50:41Z",
+        },
+        {
+            "tag_name": "b9999-draft",
+            "draft": True,
+            "prerelease": False,
+            "published_at": "2026-06-12T00:00:00Z",
+        },
+    ]
+    monkeypatch.setattr(urllib.request, "urlopen", lambda req, timeout = 5.0: _Resp(payload))
+    assert fr._fetch_latest_release_tag("unslothai/llama.cpp") == "b9596-mix-e6f2453"

studio/backend/tests/test_llama_cpp_update.py

@@ -28,23 +28,75 @@ import utils.llama_cpp_update as upd  # noqa: E402
 MARKER = "UNSLOTH_PREBUILT_INFO.json"
 
 
+class _FakeInstallerPopen:
+    """Stands in for the streamed installer process in _run_update."""
+
+    def __init__(
+        self,
+        cmd,
+        *,
+        returncode = 0,
+        lines = None,
+        on_start = None,
+        captured_kwargs = None,
+        **kwargs,
+    ):
+        if captured_kwargs is not None:
+            captured_kwargs.update(kwargs)
+        if on_start is not None:
+            on_start(list(cmd))
+        self.returncode = returncode
+        self.stdout = iter(lines or [])
+
+    def wait(self):
+        return self.returncode
+
+    def kill(self):
+        pass
+
+
+def _patch_installer_popen(
+    monkeypatch,
+    *,
+    returncode = 0,
+    lines = None,
+    on_start = None,
+    captured_kwargs = None,
+):
+    monkeypatch.setattr(
+        upd.subprocess,
+        "Popen",
+        lambda cmd, **kw: _FakeInstallerPopen(
+            cmd,
+            returncode = returncode,
+            lines = lines,
+            on_start = on_start,
+            captured_kwargs = captured_kwargs,
+            **kw,
+        ),
+    )
+
+
 def _write_install(
     dir_: Path,
     tag: str,
     repo: str = "unslothai/llama.cpp",
     asset: str | None = None,
+    release_tag: str | None = None,
 ) -> str:
     """Create a fake prebuilt install tree and return the llama-server path.
 
     ``asset`` is the bundle filename recorded in the marker; omit it to model an
-    older marker that predates asset-based ROCm forwarding (backward compat)."""
+    older marker that predates asset-based ROCm forwarding (backward compat).
+    ``release_tag`` is the full release tag (e.g. a ``b9596-mix-<sha>`` mix
+    build); defaults to ``tag`` for a plain prebuilt."""
     bin_dir = dir_ / "build" / "bin"
     bin_dir.mkdir(parents = True, exist_ok = True)
     binary = bin_dir / "llama-server"
     binary.write_text("#!/bin/sh\necho stub\n")
     marker = {
         "tag": tag,
-        "release_tag": tag,
+        "release_tag": release_tag or tag,
         "published_repo": repo,
         "installed_at_utc": "2020-01-01T00:00:00Z",
         "bundle_profile": "cuda13-newer",
@@ -57,10 +109,13 @@ def _write_install(
 
 
 @pytest.fixture(autouse = True)
-def _clean_state(monkeypatch):
+def _clean_state(monkeypatch, tmp_path):
     freshness.reset_caches()
     upd._reset_job_for_tests()
     upd._resolve_memo.clear()
+    # Isolate the freshness disk cache so the suite never writes the real
+    # ~/.unsloth cache (the default when storage_roots can't be imported).
+    monkeypatch.setattr(freshness, "_cache_dir", lambda: tmp_path / ".freshness_cache")
     # Deterministic markerless paths: no host-pinned binary, no custom dir.
     monkeypatch.delenv("LLAMA_SERVER_PATH", raising = False)
     monkeypatch.delenv("UNSLOTH_LLAMA_CPP_PATH", raising = False)
@@ -260,14 +315,15 @@ def test_start_update_source_build_installs_prebuilt(monkeypatch, tmp_path):
 
     def _fake_run(cmd, **kwargs):
         cmd = list(cmd)
-        # Status polls probe `llama-server --version`; keep the installer argv.
-        if "--version" in cmd:
-            return _Proc()
-        captured["cmd"] = cmd
-        _write_install(install_dir, "b9585")  # installer writes the marker
+        assert "--version" in cmd  # only status polls still use run()
         return _Proc()
 
+    def _on_start(cmd):
+        captured["cmd"] = cmd
+        _write_install(install_dir, "b9585")  # installer writes the marker
+
     monkeypatch.setattr(upd.subprocess, "run", _fake_run)
+    _patch_installer_popen(monkeypatch, on_start = _on_start)
 
     res = upd.start_update()
     assert res["started"] is True, res
@@ -298,21 +354,27 @@ def test_start_update_happy_path(monkeypatch, tmp_path):
         stdout = "installed"
         stderr = ""
 
-    def _fake_run(cmd, **kwargs):
-        cmd = list(cmd)
-        # Status polls probe `llama-server --version`; keep the installer argv.
-        if "--version" in cmd:
-            return _Proc()
+    def _on_start(cmd):
         captured["cmd"] = cmd
         # Simulate the installer writing a new marker with the latest tag.
         _write_install(install_dir, "b9518")
-        return _Proc()
 
-    monkeypatch.setattr(upd.subprocess, "run", _fake_run)
+    popen_kwargs: dict = {}
+    _patch_installer_popen(
+        monkeypatch,
+        lines = [
+            "[llama-prebuilt] resolving release\n",
+            "Downloading llama.zip:  35.0% (12.0 MiB/35.0 MiB) at 9.0 MiB/s\n",
+            "Downloading llama.zip:  80.0% (28.0 MiB/35.0 MiB) at 9.0 MiB/s\n",
+        ],
+        on_start = _on_start,
+        captured_kwargs = popen_kwargs,
+    )
 
     res = upd.start_update()
     assert res["started"] is True
     assert res["job"]["from_tag"] == "b9493"
+    assert res["job"]["progress"] == 0.0
 
     # Wait for the background worker.
     deadline = time.time() + 10
@@ -328,6 +390,10 @@ def test_start_update_happy_path(monkeypatch, tmp_path):
     assert str(install_dir) in captured["cmd"]
     assert "--llama-tag" in captured["cmd"] and "latest" in captured["cmd"]
     assert "unslothai/llama.cpp" in captured["cmd"]
+    # Progress lines were parsed and success pins progress at 1.0.
+    assert job["progress"] == 1.0
+    # The worker asks the installer for fine-grained progress milestones.
+    assert popen_kwargs["env"]["UNSLOTH_PROGRESS_PERCENT_STEP"] == "5"
 
 
 def test_start_update_installer_failure_reports_error(monkeypatch, tmp_path):
@@ -335,13 +401,9 @@ def test_start_update_installer_failure_reports_error(monkeypatch, tmp_path):
     binary = _write_install(install_dir, "b9493")
     monkeypatch.setattr(upd, "_find_binary", lambda: binary)
     monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
 
-    class _Proc:
-        returncode = 2
-        stdout = ""
-        stderr = "boom: network error"
-
-    monkeypatch.setattr(upd.subprocess, "run", lambda cmd, **kw: _Proc())
+    _patch_installer_popen(monkeypatch, returncode = 2, lines = ["boom: network error\n"])
 
     res = upd.start_update()
     assert res["started"] is True
@@ -410,14 +472,15 @@ def _capture_install_cmd(
 
     def _fake_run(cmd, **kwargs):
         cmd = list(cmd)
-        # Status polls probe `llama-server --version`; keep the installer argv.
-        if "--version" in cmd:
-            return _Proc()
-        captured["cmd"] = cmd
-        _write_install(install_dir, latest, repo = repo, asset = asset)
+        assert "--version" in cmd  # only status polls still use run()
         return _Proc()
 
+    def _on_start(cmd):
+        captured["cmd"] = cmd
+        _write_install(install_dir, latest, repo = repo, asset = asset)
+
     monkeypatch.setattr(upd.subprocess, "run", _fake_run)
+    _patch_installer_popen(monkeypatch, on_start = _on_start)
 
     res = upd.start_update()
     assert res["started"] is True, res
@@ -535,18 +598,12 @@ def test_update_sets_maintenance_flag_and_unloads(monkeypatch, tmp_path):
 
     seen = {}
 
-    class _Proc:
-        returncode = 0
-        stdout = "ok"
-        stderr = ""
-
-    def _fake_run(cmd, **kwargs):
+    def _on_start(cmd):
         # The maintenance flag must be set while the installer runs.
         seen["flag_during_install"] = backend._llama_update_in_progress
         _write_install(install_dir, "b9518")
-        return _Proc()
 
-    monkeypatch.setattr(upd.subprocess, "run", _fake_run)
+    _patch_installer_popen(monkeypatch, on_start = _on_start)
 
     res = upd.start_update()
     assert res["started"] is True
@@ -567,16 +624,12 @@ def test_update_clears_maintenance_flag_on_installer_failure(monkeypatch, tmp_pa
     binary = _write_install(install_dir, "b9493")
     monkeypatch.setattr(upd, "_find_binary", lambda: binary)
     monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
 
     backend = _FakeBackend()
     _inject_backend(monkeypatch, backend)
 
-    class _Proc:
-        returncode = 1
-        stdout = ""
-        stderr = "boom"
-
-    monkeypatch.setattr(upd.subprocess, "run", lambda cmd, **kw: _Proc())
+    _patch_installer_popen(monkeypatch, returncode = 1, lines = ["boom\n"])
 
     res = upd.start_update()
     assert res["started"] is True
@@ -606,16 +659,7 @@ def test_update_fails_open_when_backend_unavailable(monkeypatch, tmp_path):
     monkeypatch.setitem(sys.modules, "routes", routes_pkg)
     monkeypatch.setitem(sys.modules, "routes.inference", inference_mod)
 
-    class _Proc:
-        returncode = 0
-        stdout = "ok"
-        stderr = ""
-
-    def _fake_run(cmd, **kwargs):
-        _write_install(install_dir, "b9518")
-        return _Proc()
-
-    monkeypatch.setattr(upd.subprocess, "run", _fake_run)
+    _patch_installer_popen(monkeypatch, on_start = lambda cmd: _write_install(install_dir, "b9518"))
 
     res = upd.start_update()
     assert res["started"] is True
@@ -759,3 +803,41 @@ def test_start_update_source_build_refuses_when_newer(monkeypatch, tmp_path):
     res = upd.start_update()
     assert res["started"] is False
     assert res["reason"] == "up_to_date"
+
+
+# --- mix-tag detection + apply guard (the reported banner bug) ---
+
+
+def test_status_not_offered_on_mix_latest(monkeypatch, tmp_path):
+    # Installed the mix latest; GitHub latest is that same full tag -> no banner.
+    binary = _write_install(tmp_path / "llama.cpp", "b9596", release_tag = "b9596-mix-e6f2453")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(
+        freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9596-mix-e6f2453"
+    )
+    st = upd.get_update_status()
+    assert st["update_available"] is False
+    assert st["installed_tag"] == "b9596"
+    assert st["latest_tag"] == "b9596-mix-e6f2453"
+
+
+def test_status_not_offered_when_latest_lags(monkeypatch, tmp_path):
+    # A lagging latest (older build than installed) must never be offered.
+    binary = _write_install(tmp_path / "llama.cpp", "b9585")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9518")
+    st = upd.get_update_status()
+    assert st["update_available"] is False
+
+
+def test_start_update_marked_refuses_when_not_behind(monkeypatch, tmp_path):
+    # A direct POST / stale banner must not reinstall when already on the latest.
+    binary = _write_install(tmp_path / "llama.cpp", "b9596", release_tag = "b9596-mix-e6f2453")
+    monkeypatch.setattr(upd, "_find_binary", lambda: binary)
+    monkeypatch.setattr(upd, "_installer_script", lambda: tmp_path / "install_llama_prebuilt.py")
+    monkeypatch.setattr(
+        freshness, "_fetch_latest_release_tag", lambda repo, timeout = 5.0: "b9596-mix-e6f2453"
+    )
+    res = upd.start_update()
+    assert res["started"] is False
+    assert res["reason"] == "up_to_date"

studio/backend/tests/test_native_context_length.py

@@ -271,6 +271,7 @@ class TestPydanticModels:
     def test_load_response_has_field(self):
         """Field exists in LoadResponse.model_fields."""
         assert "native_context_length" in LoadResponse.model_fields
+        assert "context_length" in LoadResponse.model_fields
 
     def test_load_response_defaults_none(self):
         """Omitting native_context_length defaults to None."""
@@ -319,6 +320,7 @@ class TestPydanticModels:
     def test_status_response_has_field(self):
         """Field exists in InferenceStatusResponse.model_fields."""
         assert "native_context_length" in InferenceStatusResponse.model_fields
+        assert "context_length" in InferenceStatusResponse.model_fields
 
     def test_status_response_has_chat_template_field(self):
         """Status includes chat_template so the UI can rehydrate after refresh."""
@@ -347,6 +349,18 @@ class TestPydanticModels:
         roundtripped = LoadResponse.model_validate_json(resp.model_dump_json())
         assert roundtripped.native_context_length == 131072
 
+    def test_context_length_roundtrip(self):
+        """Runtime context_length serializes for non-GGUF/hub models."""
+        resp = LoadResponse(
+            status = "loaded",
+            model = "test",
+            display_name = "Test",
+            inference = {},
+            context_length = 8192,
+        )
+        roundtripped = LoadResponse.model_validate_json(resp.model_dump_json())
+        assert roundtripped.context_length == 8192
+
 
 # =====================================================================
 # D. TestRouteCompleteness -- source-level verification
@@ -408,6 +422,16 @@ class TestRouteCompleteness:
                 "native_context_length" not in block
             ), f"Non-GGUF LoadResponse should not set native_context_length:\n{block[:200]}"
 
+    def test_non_gguf_load_responses_set_runtime_context_length(self):
+        """Non-GGUF LoadResponse blocks report runtime context_length."""
+        blocks = self._find_construction_blocks("LoadResponse")
+        non_gguf = [b for b in blocks if "is_gguf = True" not in b and "is_gguf=True" not in b]
+        assert non_gguf, "Expected at least one non-GGUF LoadResponse block"
+        for block in non_gguf:
+            assert (
+                "context_length" in block
+            ), f"Non-GGUF LoadResponse should set context_length:\n{block[:200]}"
+
     def test_status_path(self):
         """InferenceStatusResponse construction with llama_backend has the field."""
         blocks = self._find_construction_blocks("InferenceStatusResponse")
@@ -420,6 +444,21 @@ class TestRouteCompleteness:
             found
         ), "No InferenceStatusResponse block with llama_backend has native_context_length"
 
+    def test_non_gguf_status_path_reports_runtime_context_length(self):
+        """Non-GGUF InferenceStatusResponse reports context_length from model_info."""
+        blocks = self._find_construction_blocks("InferenceStatusResponse")
+        found = False
+        for block in blocks:
+            if "is_gguf = False" in block and "context_length" in block:
+                found = True
+                break
+        assert found, "No non-GGUF InferenceStatusResponse block with context_length"
+
+    def test_openai_models_listing_reports_context_length(self):
+        """/v1/models includes context_length when the backend knows it."""
+        assert 'entry["context_length"]' in self._source
+        assert 'model_info.get("context_length")' in self._source
+
 
 # =====================================================================
 # E. TestEdgeCases

studio/backend/utils/llama_cpp_freshness.py

@@ -13,6 +13,7 @@ from __future__ import annotations
 
 import json
 import os
+import re
 import time
 from datetime import datetime, timezone
 from pathlib import Path
@@ -104,11 +105,18 @@ def _save_disk_cache(repo: str, latest_tag: Optional[str]) -> None:
 
 
 def _fetch_latest_release_tag(repo: str, timeout: float = 5.0) -> Optional[str]:
-    """GitHub API call. None on any failure (offline, rate-limited, etc)."""
+    """Newest published release tag for `repo`, by publish time.
+
+    Resolves "latest" the way install_llama_prebuilt.py does (newest
+    non-draft/non-prerelease by ``published_at``), NOT via GitHub's
+    ``/releases/latest`` pointer. That pointer sorts by commit date and can lag
+    behind the build the installer actually installs, so detection and apply
+    disagreed -- the cause of the downgrade/sticky banner. None on any failure
+    (offline, rate-limited, etc)."""
     import urllib.error
     import urllib.request
 
-    url = f"https://api.github.com/repos/{repo}/releases/latest"
+    url = f"https://api.github.com/repos/{repo}/releases?per_page=30"
     headers = {
         "Accept": "application/vnd.github+json",
         "User-Agent": "unsloth-studio-freshness-check",
@@ -128,8 +136,21 @@ def _fetch_latest_release_tag(repo: str, timeout: float = 5.0) -> Optional[str]:
     ) as exc:
         logger.debug("freshness fetch failed", repo = repo, error = str(exc))
         return None
-    tag = data.get("tag_name")
-    return tag if isinstance(tag, str) and tag else None
+    if not isinstance(data, list):
+        return None
+    published = [
+        r
+        for r in data
+        if isinstance(r, dict)
+        and not r.get("draft")
+        and not r.get("prerelease")
+        and isinstance(r.get("tag_name"), str)
+        and r.get("tag_name")
+    ]
+    if not published:
+        return None
+    newest = max(published, key = lambda r: r.get("published_at") or "")
+    return newest["tag_name"]
 
 
 def latest_published_release(repo: str, *, force_refresh: bool = False) -> Optional[str]:
@@ -172,19 +193,58 @@ def _parse_installed_at(value: object) -> Optional[datetime]:
     return dt
 
 
+def parse_base_build(tag: object) -> Optional[int]:
+    """Numeric base build from a release tag. Handles both a plain ``bNNNN`` and
+    a mix-build tag like ``b9596-mix-<sha>`` (anchored at the start, so the mix
+    suffix doesn't defeat it). None for anything not starting with ``bNNNN``."""
+    if not isinstance(tag, str):
+        return None
+    m = re.match(r"b(\d+)", tag.strip())
+    return int(m.group(1)) if m else None
+
+
+def is_behind(installed: Optional[str], latest: Optional[str]) -> bool:
+    """Whether `installed` is genuinely behind `latest`, comparing the FULL
+    release identity (so a mix build can legitimately be the latest) with a
+    base-build guard so a lagging GitHub /releases/latest can never read as an
+    update or a downgrade.
+
+    - identical tags -> not behind (clears the sticky banner post-update)
+    - higher base build on `latest` -> behind; lower -> NOT behind (downgrade guard)
+    - same base build: a different/new mix -> behind, but a bare ``bNNNN`` never
+      supersedes a mix build (extra PRs) at that base -> not behind
+    - non-bNNNN tags -> behind (plain inequality, since they already differ)
+    """
+    if not installed or not latest:
+        return False
+    installed, latest = installed.strip(), latest.strip()
+    if installed == latest:
+        return False
+    ib, lb = parse_base_build(installed), parse_base_build(latest)
+    if ib is None or lb is None:
+        return True
+    if lb != ib:
+        return lb > ib
+    # Same base build, different tags: offer a mix (latest carries a suffix), but
+    # never offer a bare base over a mix install at the same base.
+    return latest != f"b{lb}"
+
+
 def check_prebuilt_freshness(
     binary_path: Optional[str],
     *,
     threshold_days: int = STALENESS_THRESHOLD_DAYS,
     now: Optional[datetime] = None,
 ) -> dict:
-    """Returns {has_marker, stale, installed_tag, latest_tag,
+    """Returns {has_marker, stale, behind, installed_tag, latest_tag,
     installed_at_utc, age_days, published_repo, threshold_days}.
-    stale = True iff installed != latest AND age >= threshold.
-    Fails open on missing data (stale stays False)."""
+    behind = installed genuinely older than latest (see is_behind).
+    stale = behind AND age >= threshold.
+    Fails open on missing data (behind/stale stay False)."""
     out: dict = {
         "has_marker": False,
         "stale": False,
+        "behind": False,
         "installed_tag": None,
         "latest_tag": None,
         "installed_at_utc": None,
@@ -196,16 +256,25 @@ def check_prebuilt_freshness(
     if not marker:
         return out
     out["has_marker"] = True
+    # Display prefers the normalized base ("tag"); comparison below prefers the
+    # full "release_tag" -- deliberately opposite fallbacks.
     out["installed_tag"] = marker.get("tag") or marker.get("release_tag")
     out["installed_at_utc"] = marker.get("installed_at_utc")
     out["published_repo"] = marker.get("published_repo")
 
+    # The marker records both a normalized base tag ("tag", e.g. b9596) and the
+    # full release tag ("release_tag", e.g. b9596-mix-<sha>). Compare against the
+    # FULL identity, since GitHub /releases/latest returns the full tag_name --
+    # comparing the normalized base against the full latest is what produced the
+    # permanent "downgrade" banner on every mix release.
+    installed_full = marker.get("release_tag") or marker.get("tag")
     repo = out["published_repo"]
-    if not repo or not out["installed_tag"]:
+    if not repo or not installed_full:
         return out
     latest = latest_published_release(repo)
     out["latest_tag"] = latest
-    if not latest or latest == out["installed_tag"]:
+    out["behind"] = is_behind(installed_full, latest)
+    if not out["behind"]:
         return out
 
     installed_at = _parse_installed_at(out["installed_at_utc"])

studio/backend/utils/llama_cpp_update.py

@@ -59,10 +59,17 @@ _job: dict = {
     "from_tag": None,
     "to_tag": None,
     "error": None,
+    "progress": None,
     "started_at": None,
     "finished_at": None,
 }
 
+# Matches the installer's download progress lines, e.g.
+# "Downloading x.zip:  35.0% (12.3 MiB/35.1 MiB) at 8.2 MiB/s".
+_PROGRESS_LINE_RE = re.compile(r"(\d+(?:\.\d+)?)%\s*\(")
+# The download dominates the update; extract/validate fill the last slice.
+_DOWNLOAD_PROGRESS_CEILING = 0.95
+
 
 def _utcnow() -> str:
     return time.strftime("%Y-%m-%dT%H:%M:%SZ", time.gmtime())
@@ -279,9 +286,10 @@ def get_update_status(*, force_refresh: bool = False) -> dict:
     freshness = check_prebuilt_freshness(binary)
     installed = freshness.get("installed_tag")
     latest = freshness.get("latest_tag")
-    update_available = bool(
-        freshness.get("has_marker") and installed and latest and installed != latest
-    )
+    # `behind` compares the full release identity with a base-build guard, so a
+    # lagging /releases/latest or a mix-tagged latest can't show a false update
+    # (see llama_cpp_freshness.is_behind).
+    update_available = bool(freshness.get("has_marker") and freshness.get("behind"))
 
     with _job_lock:
         job = dict(_job)
@@ -357,19 +365,55 @@ def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path
         ]
         cmd.extend(_rocm_install_args(asset))
         logger.info("llama update: installing", cmd = " ".join(cmd))
-        proc = subprocess.run(
+        # Stream the installer output so download percent lines feed
+        # job["progress"]; finer milestones via UNSLOTH_PROGRESS_PERCENT_STEP.
+        env = dict(os.environ, UNSLOTH_PROGRESS_PERCENT_STEP = "5")
+        proc = subprocess.Popen(
             cmd,
-            capture_output = True,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.STDOUT,
             text = True,
-            timeout = _INSTALL_TIMEOUT_SECONDS,
+            env = env,
         )
-        if proc.returncode != 0:
-            tail = (proc.stderr or proc.stdout or "").strip()[-1500:]
-            raise RuntimeError(f"installer exited {proc.returncode}: {tail or 'no output'}")
+        timed_out = threading.Event()
 
-        # New UNSLOTH_PREBUILT_INFO.json is on disk; drop caches so the next
-        # status read reflects the freshly installed tag.
+        def _kill_on_timeout() -> None:
+            timed_out.set()
+            proc.kill()
+
+        watchdog = threading.Timer(_INSTALL_TIMEOUT_SECONDS, _kill_on_timeout)
+        watchdog.daemon = True
+        watchdog.start()
+        tail_lines: list[str] = []
+        try:
+            assert proc.stdout is not None
+            for line in proc.stdout:
+                tail_lines.append(line)
+                if len(tail_lines) > 80:
+                    del tail_lines[0]
+                m = _PROGRESS_LINE_RE.search(line)
+                if m is None:
+                    continue
+                fraction = min(float(m.group(1)) / 100.0, 1.0) * _DOWNLOAD_PROGRESS_CEILING
+                with _job_lock:
+                    _job["progress"] = max(_job.get("progress") or 0.0, fraction)
+            returncode = proc.wait()
+        finally:
+            watchdog.cancel()
+        if timed_out.is_set():
+            raise RuntimeError(f"installer timed out after {_INSTALL_TIMEOUT_SECONDS}s")
+        if returncode != 0:
+            tail = "".join(tail_lines).strip()[-1500:]
+            raise RuntimeError(f"installer exited {returncode}: {tail or 'no output'}")
+
+        # New UNSLOTH_PREBUILT_INFO.json is on disk; drop in-memory caches and
+        # re-prime the 24h disk freshness cache with the true newest, so the
+        # banner can't linger on a stale same-base value after the swap.
         reset_caches()
+        try:
+            latest_published_release(repo, force_refresh = True)
+        except Exception as exc:  # pragma: no cover - network defensive
+            logger.debug("llama update: post-install freshness refresh failed", error = str(exc))
         new_marker = read_install_marker(_find_binary())
         new_tag = (new_marker or {}).get("tag") or (new_marker or {}).get("release_tag")
 
@@ -382,6 +426,7 @@ def _run_update(install_dir: Path, repo: str, asset: Optional[str], script: Path
                 ),
                 to_tag = new_tag,
                 error = None,
+                progress = 1.0,
                 finished_at = _utcnow(),
             )
         logger.info("llama update: success", to_tag = new_tag)
@@ -417,7 +462,24 @@ def start_update() -> dict:
             "job": get_update_status()["job"],
         }
 
+    # A job already in flight wins over any freshness re-check below (and skips
+    # its network call). The final lock block re-checks to close the TOCTOU.
+    with _job_lock:
+        if _job["state"] == _JOB_RUNNING:
+            return {"started": False, "reason": "already_running", "job": dict(_job)}
+
     if marker:
+        # Mirror the detection guard: a direct POST or a stale banner must not
+        # start an install when the latest is not actually newer (force a fresh
+        # check so a stale 24h cache can't wrongly block a real update either).
+        status = get_update_status(force_refresh = True)
+        if not status.get("update_available"):
+            return {
+                "started": False,
+                "reason": "up_to_date",
+                "message": "The installed llama.cpp build is already at the latest prebuilt.",
+                "job": status["job"],
+            }
         install_dir = _install_dir_for(binary)
         repo = marker.get("published_repo") or DEFAULT_PUBLISHED_REPO
         from_tag = marker.get("tag") or marker.get("release_tag")
@@ -467,6 +529,7 @@ def start_update() -> dict:
             from_tag = from_tag,
             to_tag = None,
             error = None,
+            progress = 0.0,
             started_at = _utcnow(),
             finished_at = None,
         )
@@ -491,6 +554,7 @@ def _reset_job_for_tests() -> None:
             from_tag = None,
             to_tag = None,
             error = None,
+            progress = None,
             started_at = None,
             finished_at = None,
         )

studio/backend/utils/studio_version.py

@@ -15,6 +15,7 @@ _DEV_VERSION = "dev"
 _GIT_TIMEOUT_SECONDS = 1.0
 _STUDIO_TAG_RE = re.compile(r"^v\d+\.\d+\.\d+(?:-[0-9A-Za-z.][0-9A-Za-z.-]*)?$")
 _GIT_DESCRIBE_SUFFIX_RE = re.compile(r"-\d+-g[0-9A-Fa-f]+(?:-dirty)?$")
+_GIT_BRANCH_RE = re.compile(r"^[0-9A-Za-z._/-]+$")
 _MAX_VERSION_LENGTH = 64
 
 
@@ -71,6 +72,35 @@ def _exact_git_studio_tag(repo_root: Path) -> str | None:
     return tag if is_valid_studio_release_version(tag) else None
 
 
+def _git_branch(repo_root: Path) -> str | None:
+    try:
+        result = subprocess.run(
+            ["git", "rev-parse", "--abbrev-ref", "HEAD"],
+            cwd = repo_root,
+            check = False,
+            stdout = subprocess.PIPE,
+            stderr = subprocess.DEVNULL,
+            text = True,
+            timeout = _GIT_TIMEOUT_SECONDS,
+        )
+    except (OSError, subprocess.TimeoutExpired):
+        return None
+
+    if result.returncode != 0:
+        return None
+
+    branch = result.stdout.strip()
+    # "HEAD" means detached, e.g. a tag or commit checkout.
+    if (
+        not branch
+        or branch == "HEAD"
+        or len(branch) > _MAX_VERSION_LENGTH
+        or _GIT_BRANCH_RE.fullmatch(branch) is None
+    ):
+        return None
+    return branch
+
+
 def get_studio_version(repo_root: Path | None = None) -> str:
     """Return the installed Studio release tag for display, or ``dev``.
 
@@ -81,7 +111,10 @@ def get_studio_version(repo_root: Path | None = None) -> str:
 
     if _is_source_checkout(resolved_repo_root):
         git_tag = _exact_git_studio_tag(resolved_repo_root)
-        return git_tag if git_tag is not None else _DEV_VERSION
+        if git_tag is not None:
+            return git_tag
+        branch = _git_branch(resolved_repo_root)
+        return f"GitHub {branch}" if branch is not None else _DEV_VERSION
 
     stamped_version = _studio_release_build.STUDIO_RELEASE_VERSION
     if is_valid_studio_release_version(stamped_version):

studio/frontend/src/components/app-sidebar.tsx

@@ -60,6 +60,7 @@ import {
   Logout05Icon,
   MoreVerticalIcon,
   Search01Icon,
+  PlusSignIcon,
   PowerIcon,
   PencilEdit02Icon,
   LayoutAlignLeftIcon,
@@ -558,7 +559,8 @@ export function AppSidebar() {
         : "sidebar-row-action group-hover/recent-item:opacity-100 group-hover/recent-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto";
     const buttonClass = cn(
       "sidebar-nav-btn h-[33px] cursor-pointer rounded-full pr-4 text-[14.5px] leading-[19px] tracking-nav font-medium",
-      variant === "project" ? "pl-[39px]" : "pl-3",
+      // pl-3.5 starts the title at the same x as the Recents label text.
+      variant === "project" ? "pl-[39px]" : "pl-3.5",
       variant === "project"
         ? "group-hover/project-chat-item:pr-8 group-has-[.sidebar-row-action[data-state=open]]/project-chat-item:pr-8"
         : "group-hover/recent-item:pr-8 group-has-[.sidebar-row-action[data-state=open]]/recent-item:pr-8",
@@ -837,7 +839,28 @@ export function AppSidebar() {
                   navigate({ to: "/projects" });
                   closeMobileIfOpen();
                 }}
-              />
+                className="group/projects-item relative"
+              >
+                <button
+                  type="button"
+                  aria-label="New project"
+                  onClick={(e) => {
+                    e.stopPropagation();
+                    setProjectCreateMoveTarget(null);
+                    setProjectNameDraft("");
+                    setCreatingProject(true);
+                  }}
+                  className="sidebar-row-action group-hover/projects-item:opacity-100 group-hover/projects-item:pointer-events-auto focus-visible:opacity-100 focus-visible:pointer-events-auto group-data-[collapsible=icon]:hidden"
+                >
+                  <span className="sidebar-row-action-glyph">
+                    <HugeiconsIcon
+                      icon={PlusSignIcon}
+                      strokeWidth={1.75}
+                      className="size-4"
+                    />
+                  </span>
+                </button>
+              </NavItem>
               <NavItem
                 icon={DashboardCircleIcon}
                 label={t("shell.navigation.hub")}
@@ -1065,7 +1088,7 @@ export function AppSidebar() {
                       className="!size-[32px]"
                     />
                   </div>
-                  <div className="flex flex-col gap-0.5 leading-tight group-data-[collapsible=icon]:hidden">
+                  <div className="flex flex-col gap-px leading-tight group-data-[collapsible=icon]:hidden">
                     <span className="truncate font-heading text-[13.5px] tracking-[0.025em] dark:tracking-[0.04em] font-semibold text-nav-fg">{displayTitle}</span>
                     <span className="truncate text-[11.5px] tracking-nav text-muted-foreground">Unsloth</span>
                   </div>

studio/frontend/src/components/llama-update-banner.tsx

@@ -3,9 +3,10 @@
 
 import { Button } from "@/components/ui/button";
 import { useLlamaUpdateCheck } from "@/hooks/use-llama-update-check";
+import { useShowLlamaUpdateBanner } from "@/hooks/use-llama-update-pref";
 import { toast } from "@/lib/toast";
 import { AnimatePresence, motion } from "motion/react";
-import { type ReactElement, useEffect, useRef } from "react";
+import type { ReactElement } from "react";
 
 const EASE_OUT_QUART: [number, number, number, number] = [0.165, 0.84, 0.44, 1];
 
@@ -15,15 +16,19 @@ interface LlamaUpdateBannerProps {
 
 /**
  * Non-invasive "Update llama.cpp" affordance. Appears bottom-right ~1s after a
- * newer prebuilt is detected and stays up until dismissed (click outside / X)
- * or updated. Clicking Update swaps the prebuilt in place via POST /api/llama/update.
+ * newer prebuilt is detected and stays up until the user explicitly acts on it
+ * (X, Update, or Remind me later). Clicking Update swaps the prebuilt in place
+ * via POST /api/llama/update. Can be turned off entirely in Settings ->
+ * General -> Notifications (on by default).
  */
 export function LlamaUpdateBanner({
   enabled = true,
 }: LlamaUpdateBannerProps): ReactElement | null {
-  const { status, visible, applying, apply, dismiss } = useLlamaUpdateCheck({
-    enabled,
-  });
+  const showBannerPref = useShowLlamaUpdateBanner();
+  const { status, visible, applying, apply, dismiss, snooze } =
+    useLlamaUpdateCheck({
+      enabled: enabled && showBannerPref,
+    });
 
   async function handleUpdate() {
     const result = await apply();
@@ -40,30 +45,12 @@ export function LlamaUpdateBanner({
 
   const show =
     visible && status != null && (status.update_available || applying);
-  const bannerRef = useRef<HTMLDivElement>(null);
-
-  // Dismiss when the user clicks anything outside the banner. Kept off while an
-  // update is applying so the progress stays visible.
-  useEffect(() => {
-    if (!show || applying) return;
-    function onPointerDown(event: PointerEvent) {
-      if (
-        bannerRef.current &&
-        !bannerRef.current.contains(event.target as Node)
-      ) {
-        dismiss();
-      }
-    }
-    document.addEventListener("pointerdown", onPointerDown, true);
-    return () =>
-      document.removeEventListener("pointerdown", onPointerDown, true);
-  }, [show, applying, dismiss]);
+  const updateProgress = status?.job.progress ?? null;
 
   return (
     <AnimatePresence>
       {show ? (
         <motion.div
-          ref={bannerRef}
           initial={{ opacity: 0, y: 12, scale: 0.96 }}
           animate={{ opacity: 1, y: 0, scale: 1 }}
           exit={{ opacity: 0, y: 8, scale: 0.97 }}
@@ -71,12 +58,12 @@ export function LlamaUpdateBanner({
           className="fixed bottom-4 right-4 z-[9998] w-[calc(100vw-2rem)] max-w-[340px]"
           data-testid="llama-update-banner"
         >
-          <div className="corner-squircle relative overflow-hidden border border-border/60 bg-background/95 px-4 py-3 shadow-lg backdrop-blur-md">
+          <div className="relative overflow-hidden rounded-[24px] bg-white px-4 pb-[22px] pl-6 pt-5 shadow-[0_2px_8px_-2px_rgba(0,0,0,0.16)] dark:bg-card dark:shadow-[0_8px_28px_-6px_rgba(0,0,0,0.28)]">
             {applying ? null : (
               <button
                 type="button"
                 onClick={dismiss}
-                className="absolute top-2.5 right-2.5 flex size-5 items-center justify-center rounded-md text-muted-foreground/60 transition-colors hover:bg-muted hover:text-foreground"
+                className="absolute top-2.5 right-3 flex size-6 items-center justify-center rounded-full text-muted-foreground/60 transition-colors hover:bg-muted hover:text-foreground"
                 aria-label="Dismiss llama.cpp update notification"
               >
                 <svg
@@ -97,32 +84,64 @@ export function LlamaUpdateBanner({
               </button>
             )}
 
-            <div className="flex items-center gap-2 pr-5">
-              <span className="text-base" aria-hidden="true">
-                🦥
-              </span>
-              <p className="text-sm font-medium text-foreground">
-                {applying ? "Updating llama.cpp..." : "New llama.cpp prebuilt"}
+            <div className="min-w-0 pr-6">
+              <p className="font-heading text-base font-medium text-foreground">
+                {applying ? "Updating llama.cpp..." : "New llama.cpp version"}
+              </p>
+              <p className="mt-0.5 text-xs text-muted-foreground">
+                {status?.installed_tag ?? "unknown"} &rarr;{" "}
+                <span className="font-medium text-foreground">
+                  {status?.latest_tag ?? ""}
+                </span>
               </p>
             </div>
-            <p className="mt-0.5 pl-7 text-xs text-muted-foreground">
-              {status?.installed_tag ?? "unknown"} &rarr;{" "}
-              <span className="font-medium text-foreground">
-                {status?.latest_tag ?? ""}
-              </span>
-            </p>
 
-            <div className="mt-2.5 pl-7">
-              <Button
-                size="sm"
-                className="corner-squircle"
-                onClick={handleUpdate}
-                disabled={applying}
-                data-testid="llama-update-button"
+            {applying ? (
+              <div
+                className="mb-1.5 mt-4 h-1 overflow-hidden rounded-full bg-muted"
+                role="progressbar"
+                aria-label="Updating llama.cpp"
+                aria-valuemin={0}
+                aria-valuemax={100}
+                aria-valuenow={
+                  updateProgress != null
+                    ? Math.round(updateProgress * 100)
+                    : undefined
+                }
+                data-testid="llama-update-progress"
               >
-                {applying ? "Updating..." : "Update llama.cpp"}
-              </Button>
-            </div>
+                {updateProgress != null && updateProgress > 0 ? (
+                  <div
+                    className="h-full rounded-full bg-primary transition-[width] duration-700 ease-out"
+                    style={{ width: `${Math.round(updateProgress * 100)}%` }}
+                  />
+                ) : (
+                  // No percent yet (resolving the release): sweep until the
+                  // first download progress arrives.
+                  <div className="loading-bar-slide h-full w-1/3 rounded-full bg-primary" />
+                )}
+              </div>
+            ) : (
+              <div className="mt-3 flex items-center gap-2">
+                <Button
+                  size="sm"
+                  className="h-auto rounded-full px-3.5 py-2 text-[13px]"
+                  onClick={handleUpdate}
+                  data-testid="llama-update-button"
+                >
+                  Update
+                </Button>
+                <Button
+                  size="sm"
+                  variant="ghost"
+                  className="h-auto rounded-full px-2.5 py-2 text-[13px] text-muted-foreground hover:text-foreground"
+                  onClick={snooze}
+                  data-testid="llama-update-snooze-button"
+                >
+                  Remind me later
+                </Button>
+              </div>
+            )}
           </div>
         </motion.div>
       ) : null}

studio/frontend/src/components/tauri/startup-screen.tsx

@@ -228,7 +228,7 @@ function RepairingContent({
       </div>
       <div className="mb-10 flex flex-col items-center gap-2">
         <TealSpinner />
-        <p className="text-sm font-bold text-foreground">Updating existing Studio install...</p>
+        <p className="text-sm font-bold text-foreground">Updating existing Unsloth install...</p>
         {latest && (
           <p className="max-w-xs text-center text-xs text-muted-foreground">{latest}</p>
         )}

studio/frontend/src/components/ui/tooltip.tsx

@@ -98,9 +98,27 @@ function TooltipContent({
 }: React.ComponentProps<typeof TooltipPrimitive.Content> & {
   variant?: TooltipVariant;
 }) {
+  // Single-line compact tooltips render as a full pill; wrapped ones keep
+  // the squarer corners so tall pills do not look like capsules. A ref
+  // callback measures on mount: Radix mounts the portal content without
+  // re-rendering this wrapper, so an effect here would never see the node.
+  const measureRef = useCallback(
+    (el: HTMLDivElement | null) => {
+      if (!el || variant !== "default") return;
+      const cs = getComputedStyle(el);
+      const lineHeight = Number.parseFloat(cs.lineHeight) || 16;
+      const innerHeight =
+        el.clientHeight -
+        Number.parseFloat(cs.paddingTop) -
+        Number.parseFloat(cs.paddingBottom);
+      el.classList.toggle("rounded-full!", innerHeight < lineHeight * 1.5);
+    },
+    [variant],
+  );
   return (
     <TooltipPrimitive.Portal>
       <TooltipPrimitive.Content
+        ref={measureRef}
         data-slot="tooltip-content"
         sideOffset={sideOffset}
         className={cn(

studio/frontend/src/components/web/update-banner.tsx

@@ -2,13 +2,16 @@
 // Copyright 2026-present the Unsloth AI Inc. team. All rights reserved. See /studio/LICENSE.AGPL-3.0
 
 import { Button } from "@/components/ui/button";
+import { usePlatformStore } from "@/config/env";
 import { useWebUpdateCheck } from "@/hooks/use-web-update-check";
 import { isTauri } from "@/lib/api-base";
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { AnimatePresence, motion } from "motion/react";
 import { type ReactElement, useEffect, useRef, useState } from "react";
 
-const STUDIO_UPDATE_CMD = "unsloth studio update";
+const STUDIO_INSTALL_UNIX_CMD =
+  "curl -fsSL https://unsloth.ai/install.sh | sh";
+const STUDIO_INSTALL_WINDOWS_CMD = "irm https://unsloth.ai/install.ps1 | iex";
 const RELEASE_NOTES_URL = "https://unsloth.ai/docs/new/changelog";
 const EASE_OUT_QUART: [number, number, number, number] = [0.165, 0.84, 0.44, 1];
 
@@ -20,6 +23,11 @@ export function WebUpdateBanner({
   enabled = true,
 }: WebUpdateBannerProps): ReactElement | null {
   const { status, dismiss } = useWebUpdateCheck({ enabled });
+  const deviceType = usePlatformStore((s) => s.deviceType);
+  const installCmd =
+    deviceType === "windows"
+      ? STUDIO_INSTALL_WINDOWS_CMD
+      : STUDIO_INSTALL_UNIX_CMD;
   const [copiedVersion, setCopiedVersion] = useState<string | null>(null);
   const dismissTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
 
@@ -36,7 +44,7 @@ export function WebUpdateBanner({
   }
 
   async function handleCopyCommand() {
-    if (!(await copyToClipboard(STUDIO_UPDATE_CMD))) {
+    if (!(await copyToClipboard(installCmd))) {
       return;
     }
     setCopiedVersion(status?.latestVersion ?? null);
@@ -89,8 +97,8 @@ export function WebUpdateBanner({
                   Package update available: {status.latestVersion}
                 </p>
                 <p className="mt-1 text-xs leading-relaxed text-muted-foreground">
-                  Installed package: {status.currentVersion}. To update Studio,
-                  run this in your terminal, then restart Studio.
+                  Installed package: {status.currentVersion}. To update Unsloth,
+                  run this in your terminal, then restart Unsloth.
                 </p>
               </div>
             </div>

studio/frontend/src/config/env.ts

@@ -46,12 +46,28 @@ export async function fetchDeviceType(): Promise<DeviceType> {
   if (fetched) return usePlatformStore.getState().deviceType;
 
   try {
-    const res = await fetch(apiUrl("/api/health"));
+    // /api/health only reports the server's device_type to authed callers.
+    // Read the token from storage directly: importing features/auth here
+    // would be an import cycle (auth/session imports this store).
+    const token =
+      typeof window === "undefined"
+        ? null
+        : localStorage.getItem("unsloth_auth_token");
+    const res = await fetch(apiUrl("/api/health"), {
+      headers: token ? { Authorization: `Bearer ${token}` } : undefined,
+    });
     if (res.ok) {
       const data = (await res.json()) as { device_type?: string; chat_only?: boolean };
       const deviceType = data.device_type ?? detectLocalPlatform();
       const chatOnly = data.chat_only ?? false;
-      usePlatformStore.setState({ deviceType, chatOnly, fetched: true });
+      // Cache only a server-reported platform. Unauthenticated responses fall
+      // back to the browser platform, which can differ from the host (WSL,
+      // SSH); keeping fetched=false retries once a token exists.
+      usePlatformStore.setState({
+        deviceType,
+        chatOnly,
+        fetched: data.device_type !== undefined,
+      });
       return deviceType;
     }
   } catch {

studio/frontend/src/features/auth/api.ts

@@ -175,7 +175,7 @@ export async function authFetch(
           "You appear to be offline. Check your network connection and try again.",
         );
       }
-      throw new Error("Studio isn't running -- please relaunch it.");
+      throw new Error("Unsloth isn't running -- please relaunch it.");
     }
     throw err;
   }

studio/frontend/src/features/auth/tauri-auto-auth.ts

@@ -26,7 +26,7 @@ let pending: { promise: Promise<boolean>; force: boolean } | null = null;
 let lastTauriAuthFailure: string | null = null;
 
 const TAURI_AUTH_FAILURE_FALLBACK =
-  "Desktop authentication failed. Update or repair the managed Studio install, then restart Studio.";
+  "Desktop authentication failed. Update or repair the managed Unsloth install, then restart Unsloth.";
 const BACKEND_NOT_READY_MESSAGE = "Backend is not ready";
 
 function authFailureMessage(error: unknown): string {

studio/frontend/src/features/chat/adapters/studio-web-speech-dictation-adapter.ts

@@ -18,7 +18,7 @@ const describeMediaError = (error: unknown): string => {
     return "Dictation could not access the microphone.";
   }
   if (error.name === "NotAllowedError") {
-    return "Microphone access is blocked. Allow microphone access for this Studio page, then try again.";
+    return "Microphone access is blocked. Allow microphone access for this Unsloth page, then try again.";
   }
   if (error.name === "NotFoundError") {
     return "No microphone was found for dictation.";
@@ -31,7 +31,7 @@ const describeMediaError = (error: unknown): string => {
 
 const describeSpeechError = (error: string, message?: string): string => {
   if (error === "not-allowed") {
-    return "Speech recognition was blocked by the browser. Check microphone permissions for this Studio page.";
+    return "Speech recognition was blocked by the browser. Check microphone permissions for this Unsloth page.";
   }
   if (error === "service-not-allowed") {
     return "Speech recognition is blocked by the browser speech service.";

studio/frontend/src/features/chat/chat-providers-dialog.tsx

@@ -1511,7 +1511,7 @@ export function ChatProvidersSettings({
         <div className="flex min-w-0 flex-col gap-1">
           <h1 className="font-heading text-lg font-semibold">Connections</h1>
           <p className="text-xs leading-relaxed text-muted-foreground">
-            Manage model connections for chat through the Studio proxy.
+            Manage model connections for chat.
           </p>
         </div>
       </header>

studio/frontend/src/features/chat/chat-settings-sheet.tsx

@@ -325,6 +325,7 @@ function CollapsibleSection({
   label,
   labelHref,
   headerAction,
+  onLabelClick,
   children,
   defaultOpen = false,
   first = false,
@@ -342,6 +343,8 @@ function CollapsibleSection({
    * nested in a button.
    */
   headerAction?: ReactNode;
+  /** When set, clicking the label runs this instead of toggling collapse. */
+  onLabelClick?: () => void;
   children?: ReactNode;
   defaultOpen?: boolean;
   first?: boolean;
@@ -395,7 +398,7 @@ function CollapsibleSection({
         <div className={headerClasses}>
           <button
             type="button"
-            onClick={toggle}
+            onClick={onLabelClick ?? toggle}
             className="flex min-w-0 flex-1 cursor-pointer items-center text-left leading-none transition-colors hover:text-nav-fg"
           >
             <span className="leading-none">{label}</span>
@@ -558,6 +561,9 @@ export function ChatSettingsPanel({
   const [presetNameInput, setPresetNameInput] = useState(activePreset);
   const [systemPromptEditorOpen, setSystemPromptEditorOpen] = useState(false);
   const [systemPromptDraft, setSystemPromptDraft] = useState("");
+  // When the prompt overflows the inline box, clicking opens the popup editor.
+  const systemPromptBoxRef = useRef<HTMLTextAreaElement>(null);
+  const [systemPromptOverflows, setSystemPromptOverflows] = useState(false);
   const [activePresetBaseline, setActivePresetBaseline] = useState(params);
   const presets = useMemo(() => {
     return getOrderedPresets(customPresets);
@@ -762,6 +768,16 @@ export function ChatSettingsPanel({
     }
   }, [open]);
 
+  useEffect(() => {
+    const el = systemPromptBoxRef.current;
+    setSystemPromptOverflows(
+      params.systemPrompt.length > 0 &&
+        el != null &&
+        el.clientHeight > 0 &&
+        el.scrollHeight > el.clientHeight + 1,
+    );
+  }, [params.systemPrompt, open]);
+
   const settingsScrollRef = useRef<HTMLDivElement>(null);
 
   const settingsContent = (
@@ -1119,7 +1135,7 @@ export function ChatSettingsPanel({
                     />
                     <InputGroupAddon
                       align="inline-end"
-                      className="min-h-0 shrink-0 gap-0 self-stretch border-0 py-0 pl-0 !pr-1 has-[>button]:mr-0"
+                      className="min-h-0 shrink-0 gap-0 self-stretch border-0 py-0 pl-0 !pr-1 has-[>button]:mr-0 !cursor-pointer"
                     >
                       <span
                         className="!h-7 min-h-7 !w-7 min-w-7 shrink-0 self-center inline-flex items-center justify-center rounded-full border-0 px-0 text-[#a0a097] dark:text-nav-fg pointer-events-none"
@@ -1300,6 +1316,7 @@ export function ChatSettingsPanel({
         <CollapsibleSection
           label="System Prompt"
           defaultOpen={true}
+          onLabelClick={openSystemPromptEditor}
           headerAction={
             <Tooltip>
               <TooltipPrimitive.Trigger asChild>
@@ -1326,22 +1343,36 @@ export function ChatSettingsPanel({
             </Tooltip>
           }
         >
-          <button
-            type="button"
-            onClick={openSystemPromptEditor}
-            aria-label="Edit system prompt"
+          {/* Rounded wrapper clips overflowing text and the scrollbar. */}
+          <div
             className={cn(
-              "panel-text-surface -mt-1 flex w-full h-20 overflow-hidden cursor-pointer items-start px-3.5 py-2.5 text-left text-[13px] font-medium leading-relaxed corner-squircle focus-visible:outline-none focus-visible:border-ring focus-visible:ring-[1px] focus-visible:ring-ring/40",
-              params.systemPrompt
-                ? "text-nav-fg"
-                : "text-muted-foreground",
+              "panel-text-surface -mt-1 h-20 w-full overflow-hidden corner-squircle",
+              systemPromptOverflows && "cursor-pointer",
             )}
           >
-            <span className="block line-clamp-3 whitespace-pre-wrap break-words">
-              {params.systemPrompt ||
-                "Example: You are a helpful assistant..."}
-            </span>
-          </button>
+            <textarea
+              ref={systemPromptBoxRef}
+              value={params.systemPrompt}
+              onChange={(e) => set("systemPrompt")(e.target.value)}
+              onMouseDown={(e) => {
+                // Overflowing prompt: click opens the popup editor instead.
+                // While focused, clicks still move the caret normally.
+                if (
+                  systemPromptOverflows &&
+                  document.activeElement !== e.currentTarget
+                ) {
+                  e.preventDefault();
+                  openSystemPromptEditor();
+                }
+              }}
+              placeholder="Example: You are a helpful assistant..."
+              aria-label="System prompt"
+              className={cn(
+                "block size-full resize-none bg-transparent px-3.5 py-2.5 text-left text-[13px] font-medium leading-relaxed text-nav-fg outline-none placeholder:text-muted-foreground",
+                systemPromptOverflows && "cursor-pointer",
+              )}
+            />
+          </div>
         </CollapsibleSection>
 
         <CollapsibleSection label="Sampling" defaultOpen={true}>

studio/frontend/src/features/chat/tour/steps.tsx

@@ -28,7 +28,7 @@ export function buildChatTourSteps({
       body: (
         <>
           This selects what’s loaded for inference. Hub = base models. Fine-tuned
-          = trained Studio outputs, including LoRA adapters and full finetunes.
+          = trained Unsloth outputs, including LoRA adapters and full finetunes.
         </>
       ),
     },
@@ -38,7 +38,7 @@ export function buildChatTourSteps({
       title: "Two tabs",
       body: (
         <>
-          Hub: search Hugging Face models. Fine-tuned: local Studio outputs you’ve
+          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.
         </>

studio/frontend/src/features/chat/utils/chat-history-storage.ts

@@ -134,7 +134,7 @@ export function isExpectedBackgroundChatStorageError(error: unknown): boolean {
     (error.message === "Invalid or expired token" ||
       error.message === "Not authenticated" ||
       error.message === "Request failed (401)" ||
-      error.message === "Studio isn't running -- please relaunch it.")
+      error.message === "Unsloth isn't running -- please relaunch it.")
   );
 }
 

studio/frontend/src/features/hub/download-manager/api.ts

@@ -12,14 +12,14 @@ function parseErrorText(status: number, body: unknown): string {
     const detail = (body as { detail?: unknown }).detail;
     const formatted = formatFastApiDetail(detail);
     if (status === 405) {
-      return `${formatted || "Method Not Allowed"} - the Studio backend did not accept this API method. Restart Studio so the frontend and backend are on the same build.`;
+      return `${formatted || "Method Not Allowed"} - the Unsloth backend did not accept this API method. Restart Unsloth so the frontend and backend are on the same build.`;
     }
     if (formatted) return formatted;
     const message = (body as { message?: unknown }).message;
     if (typeof message === "string" && message) return message;
   }
   if (status === 405) {
-    return "Method Not Allowed - the Studio backend did not accept this API method. Restart Studio so the frontend and backend are on the same build.";
+    return "Method Not Allowed - the Unsloth backend did not accept this API method. Restart Unsloth so the frontend and backend are on the same build.";
   }
   return `Request failed (${status})`;
 }
@@ -137,7 +137,7 @@ const DOWNLOAD_TRANSPORT_CAPABILITIES_FALLBACK: DownloadTransportCapabilities =
   http: { available: true, reason: null },
   xet: {
     available: null,
-    reason: "Couldn't verify Xet support with the Studio backend.",
+    reason: "Couldn't verify Xet support with the Unsloth backend.",
   },
 };
 let downloadTransportCapabilitiesCache: DownloadTransportCapabilities | null =

studio/frontend/src/features/hub/download-manager/download-api-adapter.ts

@@ -184,7 +184,7 @@ export async function effectiveTransportMode(
     return preferred;
   }
   const reason =
-    capabilities.xet.reason ?? "Studio will use HTTP downloads instead.";
+    capabilities.xet.reason ?? "Unsloth will use HTTP downloads instead.";
   if (lastXetUnavailableWarningReason !== reason) {
     lastXetUnavailableWarningReason = reason;
     toast.warning("Xet download transport unavailable", {

studio/frontend/src/features/recipe-studio/components/executions/execution-data-tab.tsx

@@ -45,7 +45,7 @@ function formatGitHubSourceMessage(execution: RecipeExecutionRecord): string {
     return "Collecting repository threads before rows are available.";
   }
   if (source.status === "rate_limited") {
-    return source.message ?? "Waiting for GitHub rate limit. Studio will resume automatically.";
+    return source.message ?? "Waiting for GitHub rate limit. Unsloth will resume automatically.";
   }
   return source.message ?? "Collecting repository threads before rows are available.";
 }

studio/frontend/src/features/recipe-studio/components/executions/execution-overview-tab.tsx

@@ -41,7 +41,7 @@ function formatSourceMessage(execution: RecipeExecutionRecord): string {
       typeof source.retry_after_sec === "number" && source.retry_after_sec > 0
         ? ` Waiting ~${formatMetricValue(source.retry_after_sec)}s.`
         : "";
-    return `Waiting for GitHub rate limit. Studio will resume automatically.${wait}`;
+    return `Waiting for GitHub rate limit. Unsloth will resume automatically.${wait}`;
   }
   return source.message ?? "Crawling GitHub source.";
 }

studio/frontend/src/features/recipe-studio/components/runtime/execution-progress-island.tsx

@@ -68,7 +68,7 @@ function formatGitHubSourceSummary(
       typeof source.retry_after_sec === "number" && source.retry_after_sec > 0
         ? ` ~${formatMetricValue(source.retry_after_sec)}s`
         : "";
-    return `Waiting for GitHub rate limit${wait}. Studio will resume automatically.`;
+    return `Waiting for GitHub rate limit${wait}. Unsloth will resume automatically.`;
   }
   if (source.status === "retrying") {
     return source.message ?? "GitHub request failed; retrying automatically.";

studio/frontend/src/features/recipe-studio/dialogs/seed/seed-dialog.tsx

@@ -334,8 +334,8 @@ export function GithubRepoSeedForm({
         />
         <p id={tokenHelpId} className="text-xs text-muted-foreground">
           {usingEnvToken
-            ? "Studio detected a server env token, so saved/shared recipes can leave this blank."
-            : "Blank is safest for saved/shared recipes because Studio will read the server environment at run time."}
+            ? "Unsloth detected a server env token, so saved/shared recipes can leave this blank."
+            : "Blank is safest for saved/shared recipes because Unsloth will read the server environment at run time."}
         </p>
         {hasToken && (
           <p className="rounded-md bg-amber-500/10 px-2 py-1.5 text-xs text-amber-700 dark:text-amber-300">
@@ -450,7 +450,7 @@ export function GithubRepoSeedForm({
       </fieldset>
 
       <p className="text-xs text-muted-foreground">
-        Backed by Studio's built-in <code>github_repo</code> seed reader. Large
+        Backed by Unsloth's built-in <code>github_repo</code> seed reader. Large
         repos can take minutes, so start with small limits for previews.
       </p>
     </div>

studio/frontend/src/features/settings/components/update-studio-instructions.tsx

@@ -4,21 +4,29 @@
 import { copyToClipboard } from "@/lib/copy-to-clipboard";
 import { useT } from "@/i18n";
 import { cn } from "@/lib/utils";
-import { Copy01Icon, Tick02Icon } from "@hugeicons/core-free-icons";
+import {
+  ArrowUpRight01Icon,
+  Copy01Icon,
+  Tick02Icon,
+} from "@hugeicons/core-free-icons";
 import { HugeiconsIcon } from "@hugeicons/react";
 import { AnimatePresence, motion, useReducedMotion } from "motion/react";
 import type { ReactElement } from "react";
 import { useEffect, useRef, useState } from "react";
 
-const STUDIO_UPDATE_CMD = "unsloth studio update";
-const STUDIO_UPDATE_FALLBACK_UNIX_CMD =
+const STUDIO_INSTALL_UNIX_CMD =
   "curl -fsSL https://unsloth.ai/install.sh | sh";
-const STUDIO_UPDATE_FALLBACK_WINDOWS_CMD =
-  "irm https://unsloth.ai/install.ps1 | iex";
+const STUDIO_INSTALL_WINDOWS_CMD = "irm https://unsloth.ai/install.ps1 | iex";
 const STUDIO_LOCAL_PULL_CMD = "git pull --ff-only";
-const STUDIO_LOCAL_UPDATE_CMD = "unsloth studio update --local";
-const STUDIO_LOCAL_FALLBACK_UNIX_CMD = "./install.sh --local";
-const STUDIO_LOCAL_FALLBACK_WINDOWS_CMD = ".\\install.ps1 --local";
+const STUDIO_LOCAL_INSTALL_UNIX_CMD = "./install.sh --local";
+const STUDIO_LOCAL_INSTALL_WINDOWS_CMD = ".\\install.ps1 --local";
+
+const DOCS_INSTALL_URL = "https://unsloth.ai/docs/get-started/install";
+const DOCS_UPDATING_URL =
+  "https://unsloth.ai/docs/get-started/install/updating";
+const DOCS_MAC_URL = "https://unsloth.ai/docs/get-started/install/mac";
+const DOCS_WINDOWS_URL =
+  "https://unsloth.ai/docs/get-started/install/windows-installation";
 
 export type UpdateShell = "windows" | "unix";
 export type UpdateInstallSource =
@@ -30,15 +38,6 @@ export type UpdateInstallSource =
   | "unknown";
 type UpdateInstallSourceState = UpdateInstallSource | "loading";
 
-function getStudioUpdateInstructionLine(
-  shell: UpdateShell,
-  t: ReturnType<typeof useT>,
-): string {
-  return shell === "windows"
-    ? t("settings.about.update.openPowerShell")
-    : t("settings.about.update.openTerminal");
-}
-
 function isLocalInstallSource(
   installSource?: UpdateInstallSourceState | null,
 ): boolean {
@@ -128,6 +127,77 @@ function CopyableCommand({
   );
 }
 
+function DocsLink({
+  href,
+  label,
+}: {
+  href: string;
+  label: string;
+}): ReactElement {
+  return (
+    <a
+      href={href}
+      target="_blank"
+      rel="noopener noreferrer"
+      className="inline-flex items-center gap-0.5 font-medium text-muted-foreground hover:text-foreground"
+    >
+      {label}
+      <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
+    </a>
+  );
+}
+
+function UpdateDocsLinks(): ReactElement {
+  const t = useT();
+  return (
+    <p className="flex flex-wrap items-center gap-x-2 gap-y-1 text-xs text-muted-foreground leading-relaxed">
+      {t("settings.about.update.docs")}
+      <DocsLink
+        href={DOCS_INSTALL_URL}
+        label={t("settings.about.update.docsInstall")}
+      />
+      <DocsLink
+        href={DOCS_UPDATING_URL}
+        label={t("settings.about.update.docsUpdating")}
+      />
+      <DocsLink
+        href={DOCS_MAC_URL}
+        label={t("settings.about.update.docsMac")}
+      />
+      <DocsLink
+        href={DOCS_WINDOWS_URL}
+        label={t("settings.about.update.docsWindows")}
+      />
+    </p>
+  );
+}
+
+function ShellToggleButton({
+  active,
+  label,
+  onClick,
+}: {
+  active: boolean;
+  label: string;
+  onClick: () => void;
+}): ReactElement {
+  return (
+    <button
+      type="button"
+      onClick={onClick}
+      aria-pressed={active}
+      className={cn(
+        "rounded-full px-2.5 py-1 text-[11px] font-medium transition-colors",
+        active
+          ? "bg-foreground/[0.08] text-foreground dark:bg-white/[0.12]"
+          : "text-muted-foreground hover:text-foreground",
+      )}
+    >
+      {label}
+    </button>
+  );
+}
+
 // biome-ignore lint/complexity/noExcessiveCognitiveComplexity: keep source-specific update guidance in one component so the command matrix stays visible.
 export function UpdateStudioInstructions({
   className,
@@ -145,6 +215,9 @@ export function UpdateStudioInstructions({
   const shell = shellOverride ?? defaultShell;
   const prefersReducedMotion = useReducedMotion();
   const windows = shell === "windows";
+  // null means the desktop app: its bundled backend updates through the
+  // built-in updater, so terminal commands would target the wrong install.
+  const desktopManaged = installSource === null;
   const localInstallSource = isLocalInstallSource(installSource);
   const checkoutInstallSource =
     installSource === "editable" || installSource === "local_repo";
@@ -163,6 +236,22 @@ export function UpdateStudioInstructions({
     ? { opacity: 1 }
     : { opacity: 0, y: -2 };
 
+  if (desktopManaged) {
+    return (
+      <div className={cn("flex flex-col gap-3", className)}>
+        {showTitle ? (
+          <p className="shrink-0 whitespace-nowrap text-sm font-semibold font-heading">
+            {t("settings.about.update.title")}
+          </p>
+        ) : null}
+        <p className="text-xs text-muted-foreground leading-relaxed">
+          {t("settings.about.update.desktopManaged")}
+        </p>
+        <UpdateDocsLinks />
+      </div>
+    );
+  }
+
   return (
     <div className={cn("flex flex-col gap-3", className)}>
       <div
@@ -176,42 +265,55 @@ export function UpdateStudioInstructions({
             {t("settings.about.update.title")}
           </p>
         ) : null}
-        <div className="flex shrink-0 items-center gap-0.5 text-[11px]">
-          <button
-            type="button"
-            onClick={() => setShellOverride("windows")}
-            className={cn(
-              "px-0.5 py-0.5 font-medium transition-colors",
-              windows
-                ? "text-foreground"
-                : "text-muted-foreground hover:text-emerald-600",
-            )}
-            aria-pressed={windows}
-          >
-            Windows
-          </button>
-          <span className="text-border">/</span>
-          <button
-            type="button"
+        <div className="flex shrink-0 items-center gap-1.5">
+          <ShellToggleButton
+            active={!windows}
+            label="MacOS / Linux"
             onClick={() => setShellOverride("unix")}
-            className={cn(
-              "px-0.5 py-0.5 font-medium transition-colors",
-              windows
-                ? "text-muted-foreground hover:text-emerald-600"
-                : "text-foreground",
-            )}
-            aria-pressed={!windows}
-          >
-            macOS/Linux
-          </button>
+          />
+          <ShellToggleButton
+            active={windows}
+            label="Windows"
+            onClick={() => setShellOverride("windows")}
+          />
         </div>
       </div>
       {loadingInstallSource ? (
         <p className="text-xs text-muted-foreground leading-relaxed">
           {t("settings.about.update.checkingInstall")}
         </p>
-      ) : localInstallSource ? (
+      ) : (
         <>
+          <p className="text-xs text-muted-foreground leading-relaxed">
+            {t("settings.about.update.installIntro")}
+          </p>
+          <AnimatePresence mode="wait" initial={false}>
+            <motion.div
+              key={`install-${shell}`}
+              initial={fadeInitial}
+              animate={fadeAnimate}
+              exit={fadeExit}
+              transition={fadeTransition}
+            >
+              <CopyableCommand
+                command={
+                  windows ? STUDIO_INSTALL_WINDOWS_CMD : STUDIO_INSTALL_UNIX_CMD
+                }
+                copyLabel={
+                  windows
+                    ? t("settings.about.update.installCommandWindows")
+                    : t("settings.about.update.installCommandUnix")
+                }
+              />
+            </motion.div>
+          </AnimatePresence>
+        </>
+      )}
+      {loadingInstallSource ? null : localInstallSource ? (
+        <>
+          <p className="text-xs font-semibold text-foreground">
+            {t("settings.about.update.localUpdateHeading")}
+          </p>
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.localInstallDetected")}
           </p>
@@ -224,16 +326,9 @@ export function UpdateStudioInstructions({
                 command={STUDIO_LOCAL_PULL_CMD}
                 copyLabel={t("settings.about.update.gitPullCommand")}
               />
-              <CopyableCommand
-                command={STUDIO_LOCAL_UPDATE_CMD}
-                copyLabel={t("settings.about.update.localUpdateCommand")}
-              />
-              <p className="text-xs text-muted-foreground leading-relaxed">
-                {t("settings.about.update.localInstallerFallback")}
-              </p>
               <AnimatePresence mode="wait" initial={false}>
                 <motion.div
-                  key={`local-fallback-${shell}`}
+                  key={`local-installer-${shell}`}
                   initial={fadeInitial}
                   animate={fadeAnimate}
                   exit={fadeExit}
@@ -242,8 +337,8 @@ export function UpdateStudioInstructions({
                   <CopyableCommand
                     command={
                       windows
-                        ? STUDIO_LOCAL_FALLBACK_WINDOWS_CMD
-                        : STUDIO_LOCAL_FALLBACK_UNIX_CMD
+                        ? STUDIO_LOCAL_INSTALL_WINDOWS_CMD
+                        : STUDIO_LOCAL_INSTALL_UNIX_CMD
                     }
                     copyLabel={t("settings.about.update.localInstallerCommand")}
                   />
@@ -261,7 +356,7 @@ export function UpdateStudioInstructions({
               </p>
               <AnimatePresence mode="wait" initial={false}>
                 <motion.div
-                  key={`source-fallback-${shell}`}
+                  key={`source-installer-${shell}`}
                   initial={fadeInitial}
                   animate={fadeAnimate}
                   exit={fadeExit}
@@ -270,8 +365,8 @@ export function UpdateStudioInstructions({
                   <CopyableCommand
                     command={
                       windows
-                        ? STUDIO_LOCAL_FALLBACK_WINDOWS_CMD
-                        : STUDIO_LOCAL_FALLBACK_UNIX_CMD
+                        ? STUDIO_LOCAL_INSTALL_WINDOWS_CMD
+                        : STUDIO_LOCAL_INSTALL_UNIX_CMD
                     }
                     copyLabel={t("settings.about.update.localInstallerCommand")}
                   />
@@ -282,54 +377,22 @@ export function UpdateStudioInstructions({
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.restartAfterUpdate")}
           </p>
+          <UpdateDocsLinks />
         </>
       ) : unknownInstallSource ? (
         <>
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.unknownInstall")}
           </p>
-          <p className="text-xs text-muted-foreground leading-relaxed">
-            {t("settings.about.update.curlOrPypi")}
+          <p className="text-xs font-semibold text-foreground">
+            {t("settings.about.update.localUpdateHeading")}
           </p>
-          <CopyableCommand
-            command={STUDIO_UPDATE_CMD}
-            copyLabel={t("settings.about.update.updateCommand")}
-          />
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.localCheckout")}
           </p>
-          <CopyableCommand
-            command={STUDIO_LOCAL_UPDATE_CMD}
-            copyLabel={t("settings.about.update.localUpdateCommand")}
-          />
-          <p className="text-xs text-muted-foreground leading-relaxed">
-            {t("settings.about.update.restartAfterUpdate")}
-          </p>
-        </>
-      ) : (
-        <>
-          <AnimatePresence mode="wait" initial={false}>
-            <motion.p
-              key={`instruction-${shell}`}
-              initial={fadeInitial}
-              animate={fadeAnimate}
-              exit={fadeExit}
-              transition={fadeTransition}
-              className="text-xs text-muted-foreground leading-relaxed"
-            >
-              {getStudioUpdateInstructionLine(shell, t)}
-            </motion.p>
-          </AnimatePresence>
-          <CopyableCommand
-            command={STUDIO_UPDATE_CMD}
-            copyLabel={t("settings.about.update.updateCommand")}
-          />
-          <p className="text-xs text-muted-foreground leading-relaxed">
-            {t("settings.about.update.fallbackInstruction")}
-          </p>
           <AnimatePresence mode="wait" initial={false}>
             <motion.div
-              key={`fallback-${shell}`}
+              key={`local-installer-${shell}`}
               initial={fadeInitial}
               animate={fadeAnimate}
               exit={fadeExit}
@@ -338,16 +401,24 @@ export function UpdateStudioInstructions({
               <CopyableCommand
                 command={
                   windows
-                    ? STUDIO_UPDATE_FALLBACK_WINDOWS_CMD
-                    : STUDIO_UPDATE_FALLBACK_UNIX_CMD
+                    ? STUDIO_LOCAL_INSTALL_WINDOWS_CMD
+                    : STUDIO_LOCAL_INSTALL_UNIX_CMD
                 }
-                copyLabel={t("settings.about.update.fallbackCommand")}
+                copyLabel={t("settings.about.update.localInstallerCommand")}
               />
             </motion.div>
           </AnimatePresence>
           <p className="text-xs text-muted-foreground leading-relaxed">
             {t("settings.about.update.restartAfterUpdate")}
           </p>
+          <UpdateDocsLinks />
+        </>
+      ) : (
+        <>
+          <p className="text-xs text-muted-foreground leading-relaxed">
+            {t("settings.about.update.restartAfterUpdate")}
+          </p>
+          <UpdateDocsLinks />
         </>
       )}
     </div>

studio/frontend/src/features/settings/tabs/about-tab.tsx

@@ -142,7 +142,7 @@ export function AboutTab() {
         </p>
       </header>
 
-      <SettingsSection title="Studio">
+      <SettingsSection title="Unsloth">
         <SettingsRow label={t("settings.about.studioVersion")}>
           <code className="font-mono text-xs text-muted-foreground">
             {studioVersion}
@@ -207,6 +207,37 @@ export function AboutTab() {
         </SettingsRow>
       </SettingsSection>
 
+      <SettingsSection title={t("settings.about.license.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.about.license.studioLabel")}
+          description={t("settings.about.license.studioDescription")}
+        >
+          <a
+            href="https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0"
+            target="_blank"
+            rel="noopener noreferrer"
+            className="inline-flex items-center gap-1 font-mono text-xs font-medium text-muted-foreground hover:text-foreground"
+          >
+            {t("settings.about.license.studioLicense")}
+            <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
+          </a>
+        </SettingsRow>
+        <SettingsRow
+          label={t("settings.about.license.libraryLabel")}
+          description={t("settings.about.license.libraryDescription")}
+        >
+          <a
+            href="https://github.com/unslothai/unsloth/blob/main/LICENSE"
+            target="_blank"
+            rel="noopener noreferrer"
+            className="inline-flex items-center gap-1 font-mono text-xs font-medium text-muted-foreground hover:text-foreground"
+          >
+            {t("settings.about.license.libraryLicense")}
+            <HugeiconsIcon icon={ArrowUpRight01Icon} className="size-3" />
+          </a>
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection title={t("settings.about.dangerZone")}>
         <SettingsRow
           destructive={true}

studio/frontend/src/features/settings/tabs/general-tab.tsx

@@ -15,6 +15,10 @@ import { Switch } from "@/components/ui/switch";
 import { usePlatformStore } from "@/config/env";
 import { resetOnboardingDone } from "@/features/auth";
 import { useChatRuntimeStore } from "@/features/chat";
+import {
+  setShowLlamaUpdateBanner,
+  useShowLlamaUpdateBanner,
+} from "@/hooks/use-llama-update-pref";
 import {
   loadHelperPrecacheSettings,
   updateHelperPrecacheSettings,
@@ -68,6 +72,8 @@ const PREFS_KEYS: string[] = [
   "unsloth_user_profile",
   // Guided tour flags
   "tour:studio:v1",
+  // Update notifications
+  "unsloth_show_llama_update_banner",
 ];
 
 // Set by resetAllPrefs so the unmount-commit effect skips writing back the
@@ -106,6 +112,7 @@ export function GeneralTab() {
   const autoTitle = useChatRuntimeStore((s) => s.autoTitle);
   const setAutoTitle = useChatRuntimeStore((s) => s.setAutoTitle);
   const chatOnly = usePlatformStore((s) => s.chatOnly);
+  const showLlamaUpdates = useShowLlamaUpdateBanner();
   const redirectTo = `${pathname}${search}`;
 
   const [draftToken, setDraftToken] = useState(hfToken ?? "");
@@ -315,6 +322,20 @@ export function GeneralTab() {
         </SettingsRow>
       </SettingsSection>
 
+      <SettingsSection title={t("settings.general.notifications.sectionTitle")}>
+        <SettingsRow
+          label={t("settings.general.notifications.showLlamaUpdates")}
+          description={t(
+            "settings.general.notifications.showLlamaUpdatesDescription",
+          )}
+        >
+          <Switch
+            checked={showLlamaUpdates}
+            onCheckedChange={setShowLlamaUpdateBanner}
+          />
+        </SettingsRow>
+      </SettingsSection>
+
       <SettingsSection title={t("settings.general.uploads.sectionTitle")}>
         <SettingsRow
           label={t("settings.general.uploads.maxUploadSize")}

studio/frontend/src/features/studio/tour/steps/nav.tsx

@@ -9,7 +9,7 @@ export const studioNavStep: TourStep = {
   title: "Quick orientation",
   body: (
     <>
-      Studio: pick base model, dataset, hyperparams, then start training. After
+      Unsloth: pick base model, dataset, hyperparams, then start training. After
       you start, you’ll see a Training view with live loss/metrics. Chat is for
       testing base vs LoRA adapters. Export packages checkpoints for deployment.{" "}
       <ReadMore href="https://unsloth.ai/docs/get-started/fine-tuning-for-beginners" />

studio/frontend/src/hooks/use-llama-update-check.ts

@@ -5,11 +5,14 @@ import { authFetch, getAuthToken } from "@/features/auth";
 import { useCallback, useEffect, useRef, useState } from "react";
 
 // First check shortly after load, then re-surface as an hourly reminder. The
-// banner stays up until the user dismisses it (click outside / X) or updates.
+// banner stays up until the user explicitly acts on it (X, Update, or
+// Remind me later).
 const FIRST_CHECK_DELAY_MS = 1000;
 const REMINDER_INTERVAL_MS = 60 * 60 * 1000; // ~1 hour
+// "Remind me later" re-surfaces sooner than the hourly reminder.
+const SNOOZE_DELAY_MS = 15 * 60 * 1000; // ~15 minutes
 // While an update is applying, poll the job state at this cadence.
-const JOB_POLL_INTERVAL_MS = 3000;
+const JOB_POLL_INTERVAL_MS = 1500;
 
 export interface LlamaUpdateJob {
   state: "idle" | "running" | "success" | "error";
@@ -17,6 +20,8 @@ export interface LlamaUpdateJob {
   from_tag: string | null;
   to_tag: string | null;
   error: string | null;
+  // Download fraction (0..1) while running, 1 on success, null when unknown.
+  progress: number | null;
 }
 
 export interface LlamaUpdateStatus {
@@ -42,6 +47,7 @@ function parseStatus(value: unknown): LlamaUpdateStatus | null {
       from_tag: typeof job.from_tag === "string" ? job.from_tag : null,
       to_tag: typeof job.to_tag === "string" ? job.to_tag : null,
       error: typeof job.error === "string" ? job.error : null,
+      progress: typeof job.progress === "number" ? job.progress : null,
     },
   };
 }
@@ -73,9 +79,10 @@ export interface LlamaApplyResult {
 
 /**
  * Polls the backend for a newer llama.cpp prebuilt. When one exists, `visible`
- * becomes true ~1s after load and stays up until the user dismisses it (click
- * outside / X) or updates; it re-surfaces every ~hour as a reminder. `apply()`
- * triggers the in-place swap and tracks the job.
+ * becomes true ~1s after load and stays up until the user dismisses it (X),
+ * snoozes it ("Remind me later", ~15 min), or updates; it re-surfaces every
+ * ~hour as a reminder. `apply()` triggers the in-place swap and tracks the
+ * job.
  */
 export function useLlamaUpdateCheck({
   enabled = true,
@@ -84,6 +91,7 @@ export function useLlamaUpdateCheck({
   const [visible, setVisible] = useState(false);
   const [applying, setApplying] = useState(false);
   const pollTimer = useRef<ReturnType<typeof setInterval> | null>(null);
+  const snoozeTimer = useRef<ReturnType<typeof setTimeout> | null>(null);
 
   const clearPollTimer = useCallback(() => {
     if (pollTimer.current) {
@@ -161,6 +169,10 @@ export function useLlamaUpdateCheck({
       clearTimeout(firstTimer);
       clearInterval(reminder);
       clearPollTimer();
+      if (snoozeTimer.current) {
+        clearTimeout(snoozeTimer.current);
+        snoozeTimer.current = null;
+      }
     };
   }, [enabled, surfaceIfAvailable, clearPollTimer]);
 
@@ -168,6 +180,16 @@ export function useLlamaUpdateCheck({
     setVisible(false);
   }, []);
 
+  // Hide now, re-check and re-surface after SNOOZE_DELAY_MS.
+  const snooze = useCallback(() => {
+    setVisible(false);
+    if (snoozeTimer.current) clearTimeout(snoozeTimer.current);
+    snoozeTimer.current = setTimeout(() => {
+      snoozeTimer.current = null;
+      fetchStatus().then(surfaceIfAvailable);
+    }, SNOOZE_DELAY_MS);
+  }, [surfaceIfAvailable]);
+
   const apply = useCallback(async (): Promise<LlamaApplyResult> => {
     if (applying) return { ok: false, error: "already running" };
     setApplying(true);
@@ -219,5 +241,6 @@ export function useLlamaUpdateCheck({
     applying,
     apply,
     dismiss,
+    snooze,
   };
 }

studio/frontend/src/hooks/use-llama-update-pref.ts

@@ -0,0 +1,49 @@
+// 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 { useSyncExternalStore } from "react";
+
+// Whether the llama.cpp update banner may appear. On by default; only an
+// explicit "false" (Settings -> General -> Notifications) disables it.
+const STORAGE_KEY = "unsloth_show_llama_update_banner";
+
+const listeners = new Set<() => void>();
+
+export function getShowLlamaUpdateBanner(): boolean {
+  try {
+    return localStorage.getItem(STORAGE_KEY) !== "false";
+  } catch {
+    return true;
+  }
+}
+
+export function setShowLlamaUpdateBanner(show: boolean): void {
+  try {
+    if (show) {
+      // Remove rather than store "true" so the default stays on.
+      localStorage.removeItem(STORAGE_KEY);
+    } else {
+      localStorage.setItem(STORAGE_KEY, "false");
+    }
+  } catch {
+    // storage unavailable
+  }
+  for (const listener of listeners) listener();
+}
+
+function subscribe(listener: () => void): () => void {
+  listeners.add(listener);
+  // Sync toggles made in another tab.
+  const onStorage = (event: StorageEvent) => {
+    if (event.key === STORAGE_KEY) listener();
+  };
+  window.addEventListener("storage", onStorage);
+  return () => {
+    listeners.delete(listener);
+    window.removeEventListener("storage", onStorage);
+  };
+}
+
+export function useShowLlamaUpdateBanner(): boolean {
+  return useSyncExternalStore(subscribe, getShowLlamaUpdateBanner);
+}

studio/frontend/src/hooks/use-tauri-backend.ts

@@ -56,23 +56,23 @@ function wait(ms: number) {
 function externalConflictMessage(preflight: DesktopPreflightResult) {
   if (preflight.reason === "desktop_owned_backend_active") {
     return preflight.port
-      ? `A desktop-owned Studio server for this install is already running on port ${preflight.port}. Quit the other desktop app instance, then try again.`
-      : "A desktop-owned Studio server for this install is already running. Quit the other desktop app instance, then try again.";
+      ? `A desktop-owned Unsloth server for this install is already running on port ${preflight.port}. Quit the other desktop app instance, then try again.`
+      : "A desktop-owned Unsloth server for this install is already running. Quit the other desktop app instance, then try again.";
   }
 
   if (preflight.reason === "desktop_owned_backend_starting") {
-    return "The desktop-owned Studio backend is still starting. Wait a moment, then try again.";
+    return "The desktop-owned Unsloth backend is still starting. Wait a moment, then try again.";
   }
 
   if (preflight.reason?.startsWith("desktop_owned_backend_unmanageable:")) {
     return preflight.port
-      ? `A desktop-owned Studio backend on port ${preflight.port} cannot be safely controlled by this desktop app. Stop that backend, then reopen Studio.`
-      : "A desktop-owned Studio backend cannot be safely controlled by this desktop app. Stop that backend, then reopen Studio.";
+      ? `A desktop-owned Unsloth backend on port ${preflight.port} cannot be safely controlled by this desktop app. Stop that backend, then reopen Unsloth.`
+      : "A desktop-owned Unsloth backend cannot be safely controlled by this desktop app. Stop that backend, then reopen Unsloth.";
   }
 
   return preflight.port
-    ? `A Studio server for this install is already running from a terminal on port ${preflight.port}. Stop that server, or run \`unsloth studio update\` from that terminal before using the desktop app.`
-    : "A Studio server for this install is already running from a terminal. Stop that server, or run `unsloth studio update` from that terminal before using the desktop app.";
+    ? `A Unsloth server for this install is already running from a terminal on port ${preflight.port}. Stop that server, or run \`unsloth studio update\` from that terminal before using the desktop app.`
+    : "A Unsloth server for this install is already running from a terminal. Stop that server, or run `unsloth studio update` from that terminal before using the desktop app.";
 }
 
 async function waitForManagedServerReady(
@@ -248,8 +248,8 @@ export function useTauriBackend() {
           } else {
             setBackendError(
               preflight.disposition === "owned_stale"
-                ? "Desktop-owned Studio backend is too old for this desktop app. Run `unsloth studio update`, then restart Studio."
-                : "Managed Studio install is too old. Run `unsloth studio update`.",
+                ? "Desktop-owned Unsloth backend is too old for this desktop app. Run `unsloth studio update`, then restart Unsloth."
+                : "Managed Unsloth install is too old. Run `unsloth studio update`.",
             );
           }
           return;
@@ -304,7 +304,7 @@ export function useTauriBackend() {
       if (msg.includes("already running")) {
         startingRef.current = false;
         setBackendError(
-          "Managed server is already running but did not report a port. Restart Studio and try again.",
+          "Managed server is already running but did not report a port. Restart Unsloth and try again.",
         );
         return;
       }
@@ -596,7 +596,7 @@ export function useTauriBackend() {
       const detail =
         event instanceof CustomEvent && typeof event.detail === "string"
           ? event.detail
-          : "Desktop authentication failed. Update or repair the managed Studio install, then restart Studio.";
+          : "Desktop authentication failed. Update or repair the managed Unsloth install, then restart Unsloth.";
       setAuthFailure(detail);
     };
     window.addEventListener("tauri-auth-failed", onAuthFailed);

studio/frontend/src/i18n/locales/en.ts

@@ -83,7 +83,7 @@ export const en = {
     title: "Settings",
     dialog: {
       title: "Settings",
-      description: "Manage your Unsloth Studio preferences.",
+      description: "Manage your Unsloth preferences.",
       closeAriaLabel: "Close settings",
     },
     tabs: {
@@ -93,11 +93,11 @@ export const en = {
       chat: "Chat",
       connections: "Connections",
       apiKeys: "API",
-      about: "Help",
+      about: "About",
     },
     general: {
       title: "General",
-      description: "Global preferences for Unsloth Studio.",
+      description: "Global preferences for Unsloth.",
       account: "Account",
       huggingFaceToken: "Hugging Face token",
       huggingFaceTokenDescription:
@@ -112,38 +112,44 @@ export const en = {
         sectionTitle: "Helper LLM",
         preloadOnStartup: "Pre-cache Helper LLM on startup",
         preloadOnStartupDescription:
-          "Download and cache the AI Assist helper model in the background when Studio starts. Off by default; AI Assist can still download it on demand when clicked.",
+          "Download the AI Assist helper model in the background on startup. Off by default; AI Assist can still fetch it on demand.",
         disabledByEnv:
           "Disabled by UNSLOTH_HELPER_MODEL_DISABLE in the backend environment.",
         loadError: "Failed to load Helper LLM settings.",
         saveError: "Failed to save Helper LLM settings.",
       },
+      notifications: {
+        sectionTitle: "Notifications",
+        showLlamaUpdates: "llama.cpp update notifications",
+        showLlamaUpdatesDescription:
+          "Notify when a newer llama.cpp build is available. Turn off if you only train.",
+      },
       gettingStarted: "Getting started",
       startOnboarding: "Start onboarding",
       startOnboardingDescription:
-        "Open the setup wizard again without changing your account.",
+        "Reopen the setup wizard without changing your account.",
       startOnboardingAction: "Start onboarding",
       uploads: {
         sectionTitle: "Uploads",
         maxUploadSize: "Training dataset upload cap",
         maxUploadSizeDescription:
-          "Applies to training dataset uploads. Default is {defaultSize} MB.",
+          "Default is {defaultSize} MB.",
       },
       resetPreferences: {
         sectionTitle: "Danger zone",
         label: "Reset all local preferences",
         description:
-          "Clears local-only preferences. Chats, API access, and DB-backed settings are not affected.",
+          "Clears local-only preferences. Chats, API access, and DB-backed settings are kept.",
         action: "Reset preferences",
         confirmTitle: "Reset all local preferences?",
         confirmDescription:
-          "This clears local-only preferences, then reloads Studio. Chats, API access, and DB-backed settings are not affected.",
+          "Clears local-only preferences and reloads Unsloth. Chats, API access, and DB-backed settings are kept.",
         confirmAction: "Reset and reload",
       },
     },
     profile: {
       title: "Profile",
-      description: "Update how your profile appears in Studio.",
+      description: "How your profile appears in Unsloth.",
       changePicture: "Change profile picture",
       displayName: "Display name",
       nameSaved: "Profile name saved",
@@ -163,7 +169,7 @@ export const en = {
       theme: {
         title: "Theme",
         label: "Color scheme",
-        description: "Choose light, dark, or follow your system.",
+        description: "Light, dark, or follow your system.",
         system: "System",
         light: "Light",
         dark: "Dark",
@@ -171,7 +177,7 @@ export const en = {
       language: {
         title: "Language",
         label: "Display language",
-        description: "Choose the language used by Studio.",
+        description: "The language used by Unsloth.",
       },
       layout: {
         title: "Layout",
@@ -182,25 +188,25 @@ export const en = {
     },
     chat: {
       title: "Chat",
-      description: "Manage your chat history stored on this device.",
+      description: "Manage chat history stored on this device.",
       artifacts: {
         title: "Artifacts",
         collapseHtmlBlocks: "Collapse HTML blocks",
         collapseHtmlBlocksDescription:
-          "Artifacts mode collapses full HTML fallback automatically. Turn this on to also collapse full fenced HTML documents when Artifacts is off.",
+          "Artifacts mode collapses full HTML automatically. Turn on to also collapse fenced HTML documents when Artifacts is off.",
         allowNetworkAccess: "Allow artifact network access",
         allowNetworkAccessDescription:
-          "Let artifact previews load scripts, styles, fonts, media, fetch, and WebSocket resources from HTTP(S) CDNs. Keep off for fully offline previews.",
+          "Let artifact previews load scripts, styles, fonts, media, and network resources from CDNs. Keep off for fully offline previews.",
       },
       data: "Data",
       exportHistory: "Export chat history",
       exportHistoryDescription:
-        "Download all chats and messages as a JSON file.",
+        "Download all chats and messages as JSON.",
       exportAction: "Export",
       exportingAction: "Exporting...",
       exportConversations: "Export Recents and Projects",
       exportConversationsDescription:
-        "Download Recents only, or Recents plus project chats, as Raw JSONL, CSV, or ShareGPT JSONL, combined or one file per chat.",
+        "Download Recents or Recents plus project chats as Raw JSONL, CSV, or ShareGPT JSONL, combined or per chat.",
       exportConversationsAction: "Export",
       exportScopeRecents: "Recents",
       exportScopeAll: "Recents + Projects",
@@ -208,14 +214,14 @@ export const en = {
       exportPerChatSuffix: "(per chat)",
       importChats: "Import chats",
       importChatsDescription:
-        "Add conversations from a JSONL, NDJSON, or CSV export to Recents.",
+        "Import a JSONL, NDJSON, or CSV export into Recents.",
       importChatsAction: "Import",
       importNoConversations: "No conversations found in file.",
       importedOneChat: "Imported 1 conversation to Recents.",
       importedChatCount: "Imported {count} conversations to Recents.",
       importFailed: "Import failed.",
       clearHistory: "Clear chat history",
-      clearHistoryDescription: "Delete local chat history from this device.",
+      clearHistoryDescription: "Delete chat history from this device.",
       clearAction: "Clear",
       clearAllChats: "Clear all chats",
       clearAllChatsDescription: "Permanently delete every chat on this device.",
@@ -228,7 +234,7 @@ export const en = {
       clearOneChatTitle: "Clear 1 chat?",
       clearChatsTitle: "Clear {count} chats?",
       clearChatsConfirmDescription:
-        "This permanently deletes every chat and message stored on this device. This cannot be undone.",
+        "Permanently deletes every chat on this device. This cannot be undone.",
       clearingAction: "Clearing...",
       clearOneChatAction: "Clear 1 chat",
       clearChatCountAction: "Clear {count} chats",
@@ -251,12 +257,12 @@ export const en = {
     },
     connections: {
       title: "Connections",
-      description: "Manage providers and external service connections.",
+      description: "Manage providers and external connections.",
     },
     apiKeys: {
       title: "API",
       description:
-        "Access Unsloth programmatically via the OpenAI-compatible API.",
+        "Access Unsloth via the OpenAI-compatible API.",
       readDocs: "Read the API docs",
       noAccess: "No API access yet.",
       newBadge: "New",
@@ -296,62 +302,71 @@ export const en = {
       revokeToken: "Revoke token",
       revokeTitle: 'Revoke access token "{name}"?',
       revokeDescription:
-        "Applications using this token will immediately lose access. This cannot be undone.",
+        "Apps using this token immediately lose access. This cannot be undone.",
       revokeAction: 'Revoke "{name}"',
       revoking: "Revoking...",
     },
     about: {
       title: "About",
       description:
-        "Documentation, release notes, feedback, and Studio build info.",
-      studioVersion: "Studio Version",
+        "Docs, release notes, feedback, and build info.",
+      studioVersion: "Unsloth Version",
       packageVersion: "Package Version",
-      updates: "Updates",
+      updates: "Update",
       help: "Help",
       documentation: "Documentation",
       releaseNotes: "Release notes",
       whatsNew: "What's new",
       feedback: "Feedback",
       reportIssue: "Report an issue",
+      license: {
+        sectionTitle: "License",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription:
+          "Open source under the GNU AGPL v3.0.",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "Licensed under Apache 2.0.",
+      },
       dangerZone: "Danger zone",
       shutDownStudio: "Shut down Unsloth Studio",
       shutDownStudioDescription:
-        "Stops the Studio server process and ends your session.",
+        "Stops the Unsloth server and ends your session.",
       shutDown: "Shut down",
       update: {
         title: "Update Unsloth Studio",
-        openPowerShell: "Open PowerShell and run:",
-        openTerminal: "Open Terminal and run:",
         commandText: "{label} text",
         copied: "Copied",
         copyCommand: "Copy command",
         commandCopied: "{label} copied",
         copyNamedCommand: "Copy {label}",
-        checkingInstall: "Checking how Studio was installed...",
+        checkingInstall: "Checking how Unsloth was installed...",
+        installIntro: "To install or update Unsloth:",
+        localUpdateHeading: "Local update",
+        installCommandUnix: "macOS/Linux install command",
+        installCommandWindows: "Windows install command",
         localInstallDetected:
-          "Source or local install detected. To avoid replacing it with PyPI, update from the checkout or source you originally installed from.",
-        pullThenUpdate:
-          "Pull latest changes from your Unsloth repo checkout, then update Studio locally:",
+          "Local install detected. Update from your original checkout to avoid replacing it with PyPI.",
+        pullThenUpdate: "Pull the latest changes, then run the local installer:",
         gitPullCommand: "git pull command",
-        localUpdateCommand: "local update command",
-        localInstallerFallback:
-          "If the Studio update command is unavailable, run the local installer from that checkout:",
         localInstallerCommand: "local installer command",
         sourceInstallDetected:
-          "This looks like a source or VCS package install. Reinstall from the original local path or Git URL you used.",
+          "Source or VCS package install detected. Reinstall from the original local path or Git URL.",
         repoCheckoutFallback:
-          "If you still have the Unsloth repo checkout, run the local installer from that checkout:",
-        restartAfterUpdate:
-          "Restart Studio after updating for changes to take effect.",
+          "If you still have the repo checkout, run the local installer from it:",
+        restartAfterUpdate: "Restart Unsloth after updating.",
+        desktopManaged:
+          "The desktop app keeps its bundled backend updated and will prompt when a new version is available.",
         unknownInstall:
-          "Studio could not detect how it was installed. Check how you installed Studio first, then choose the matching update path.",
-        curlOrPypi: "For curl or PyPI installs, run:",
-        updateCommand: "update command",
+          "Could not detect how Unsloth was installed. For installer or PyPI installs, use the commands above.",
         localCheckout:
-          "For local checkout installs, update from that checkout instead and use the local update command:",
-        fallbackInstruction:
-          "If that fails or unsloth studio update is unavailable, run:",
-        fallbackCommand: "fallback command",
+          "For local checkout installs, run the local installer from that checkout:",
+        docs: "Install docs:",
+        docsInstall: "Installation",
+        docsUpdating: "Updating",
+        docsMac: "Mac",
+        docsWindows: "Windows",
       },
     },
   },

studio/frontend/src/i18n/locales/zh-CN.ts

@@ -93,7 +93,7 @@ export const zhCN = {
       chat: "聊天",
       connections: "连接",
       apiKeys: "API",
-      about: "帮助",
+      about: "关于",
     },
     general: {
       title: "通用",
@@ -106,6 +106,12 @@ export const zhCN = {
       chatDefaults: "聊天默认设置",
       autoTitleNewChats: "自动为新聊天命名",
       autoTitleNewChatsDescription: "根据第一条消息生成简短标题。",
+      notifications: {
+        sectionTitle: "通知",
+        showLlamaUpdates: "llama.cpp 更新通知",
+        showLlamaUpdatesDescription:
+          "有新的 llama.cpp 构建时提醒。如果只用于训练可以关闭。",
+      },
       gettingStarted: "入门",
       startOnboarding: "开始引导",
       startOnboardingDescription: "重新打开设置向导，不会更改你的账号。",
@@ -124,13 +130,13 @@ export const zhCN = {
         action: "重置偏好设置",
         confirmTitle: "重置所有本地偏好设置？",
         confirmDescription:
-          "这会清除仅保存在本地的偏好设置，然后重新加载 Studio。聊天、API 访问权限和数据库中的设置不会受到影响。",
+          "这会清除仅保存在本地的偏好设置，然后重新加载 Unsloth。聊天、API 访问权限和数据库中的设置不会受到影响。",
         confirmAction: "重置并重新加载",
       },
     },
     profile: {
       title: "个人资料",
-      description: "更新你在 Studio 中显示的个人资料。",
+      description: "更新你在 Unsloth 中显示的个人资料。",
       changePicture: "更换头像",
       displayName: "显示名称",
       nameSaved: "个人资料名称已保存",
@@ -150,7 +156,7 @@ export const zhCN = {
       language: {
         title: "语言",
         label: "显示语言",
-        description: "选择 Studio 使用的语言。",
+        description: "选择 Unsloth 使用的语言。",
       },
       theme: {
         title: "主题",
@@ -274,9 +280,9 @@ export const zhCN = {
       revoking: "撤销中...",
     },
     about: {
-      title: "帮助",
-      description: "文档、发布说明、反馈和 Studio 构建信息。",
-      studioVersion: "Studio 版本",
+      title: "关于",
+      description: "文档、发布说明、反馈和 Unsloth 构建信息。",
+      studioVersion: "Unsloth 版本",
       packageVersion: "包版本",
       updates: "更新",
       help: "帮助",
@@ -285,43 +291,52 @@ export const zhCN = {
       whatsNew: "最新内容",
       feedback: "反馈",
       reportIssue: "报告问题",
+      license: {
+        sectionTitle: "许可证",
+        studioLabel: "Unsloth Studio",
+        studioLicense: "AGPL-3.0",
+        studioDescription: "基于 GNU AGPL v3.0 开源。",
+        libraryLabel: "Unsloth Core",
+        libraryLicense: "Apache-2.0",
+        libraryDescription: "基于 Apache License 2.0 许可。",
+      },
       dangerZone: "危险区域",
       shutDownStudio: "关闭 Unsloth Studio",
-      shutDownStudioDescription: "停止 Studio 服务进程并结束你的会话。",
+      shutDownStudioDescription: "停止 Unsloth 服务进程并结束你的会话。",
       shutDown: "关闭",
       update: {
         title: "更新 Unsloth Studio",
-        openPowerShell: "打开 PowerShell 并运行：",
-        openTerminal: "打开终端并运行：",
         commandText: "{label} 文本",
         copied: "已复制",
         copyCommand: "复制命令",
         commandCopied: "{label} 已复制",
         copyNamedCommand: "复制 {label}",
-        checkingInstall: "正在检查 Studio 的安装方式...",
+        checkingInstall: "正在检查 Unsloth 的安装方式...",
+        installIntro: "安装或更新 Unsloth：",
+        localUpdateHeading: "本地更新",
+        installCommandUnix: "macOS/Linux 安装命令",
+        installCommandWindows: "Windows 安装命令",
         localInstallDetected:
-          "检测到源码或本地安装。为避免替换为 PyPI 版本，请从最初安装时使用的 checkout 或源码位置更新。",
-        pullThenUpdate:
-          "从你的 Unsloth 仓库 checkout 拉取最新变更，然后本地更新 Studio：",
+          "检测到本地安装。请从最初的 checkout 更新，以免被 PyPI 版本替换。",
+        pullThenUpdate: "拉取最新变更，然后运行本地安装器：",
         gitPullCommand: "git pull 命令",
-        localUpdateCommand: "本地更新命令",
-        localInstallerFallback:
-          "如果 Studio 更新命令不可用，请从该 checkout 运行本地安装器：",
         localInstallerCommand: "本地安装器命令",
         sourceInstallDetected:
           "这看起来是源码或 VCS 包安装。请从最初使用的本地路径或 Git URL 重新安装。",
         repoCheckoutFallback:
           "如果你仍保留 Unsloth 仓库 checkout，请从该 checkout 运行本地安装器：",
-        restartAfterUpdate: "更新后重启 Studio，使变更生效。",
+        restartAfterUpdate: "更新后请重启 Unsloth。",
+        desktopManaged:
+          "桌面应用会自动更新其内置后端，有新版本时会提示。",
         unknownInstall:
-          "Studio 无法检测安装方式。请先确认你如何安装 Studio，然后选择匹配的更新方式。",
-        curlOrPypi: "对于 curl 或 PyPI 安装，请运行：",
-        updateCommand: "更新命令",
+          "Unsloth 无法检测安装方式。如果你使用一键安装器或 PyPI 安装，请使用上面的命令。",
         localCheckout:
-          "对于本地 checkout 安装，请改为从该 checkout 更新并使用本地更新命令：",
-        fallbackInstruction:
-          "如果失败，或 unsloth studio update 不可用，请运行：",
-        fallbackCommand: "备用命令",
+          "对于本地 checkout 安装，请改为从该 checkout 运行本地安装器：",
+        docs: "安装文档：",
+        docsInstall: "安装",
+        docsUpdating: "更新",
+        docsMac: "Mac",
+        docsWindows: "Windows",
       },
     },
   },

studio/frontend/src/index.css

@@ -552,6 +552,7 @@
 	.sidebar-nav-btn[data-state="open"],
 	.group\/project-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
 	.group\/project-chat-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.group\/projects-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
 	.group\/recent-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
 	.group\/run-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn {
 		background-color: var(--nav-surface-hover) !important;
@@ -562,6 +563,7 @@
 	.dark .sidebar-nav-btn[data-state="open"],
 	.dark .group\/project-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
 	.dark .group\/project-chat-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
+	.dark .group\/projects-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
 	.dark .group\/recent-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn,
 	.dark .group\/run-item:hover:not(:has(.sidebar-row-action:hover)) .sidebar-nav-btn {
 		color: #fff !important;
@@ -957,6 +959,11 @@
 		border-radius: 14px !important;
 	}
 
+	/* Account menu: a touch rounder than standard list menus. */
+	[data-slot="dropdown-menu-content"].app-user-menu.menu-soft-surface-up {
+		border-radius: 18px !important;
+	}
+
 	/* Every dropdown/menu/select/popover: borderless, chatbox shadow in light,
 	   none in dark. !important so it also overrides the bespoke menu shadows. */
 	[data-slot="dropdown-menu-content"],
@@ -1078,13 +1085,16 @@
 	[data-pill-compact="true"]
 		.composer-pill-btn:not([data-keep-label])[data-pill-label]:hover::after {
 		content: attr(data-pill-label);
-		@apply pointer-events-none absolute bottom-[calc(100%+6px)] left-1/2 z-50 -translate-x-1/2 rounded-[9px] bg-black px-2.5 py-1.5 text-[11px] font-medium leading-snug whitespace-nowrap text-white shadow-md;
+		/* Always one nowrap line, so always a full pill. */
+		@apply pointer-events-none absolute bottom-[calc(100%+6px)] left-1/2 z-50 -translate-x-1/2 rounded-full bg-black px-2.5 py-1.5 text-[11px] font-medium leading-snug whitespace-nowrap text-white shadow-md;
 	}
 
 	/* Compact caret pills (RAG, MCP) open their menu on click instead of
-	   toggling off, so keep the icon and skip the X swap. */
+	   toggling off, so keep the icon and skip the X swap. No data-active
+	   requirement: the off-switch glyph rules below hide the icon on hover
+	   even for inactive pills, which left a blank slot in compact mode. */
 	[data-pill-compact="true"]
-		.composer-pill-btn:not([data-keep-label]):has(.composer-pill-caret)[data-active="true"]:hover
+		.composer-pill-btn:not([data-keep-label]):has(.composer-pill-caret):hover
 		.composer-pill-glyph
 		> :not(.composer-pill-x) {
 		opacity: 1;
@@ -1512,6 +1522,20 @@
 		}
 	}
 
+	/* Indeterminate loading bar: a 1/3-width segment sweeping the track. */
+	.loading-bar-slide {
+		animation: loading-bar-slide 1.3s ease-in-out infinite;
+	}
+
+	@keyframes loading-bar-slide {
+		from {
+			transform: translateX(-110%);
+		}
+		to {
+			transform: translateX(420%);
+		}
+	}
+
 	@keyframes artifact-loading-line {
 		0% {
 			transform: translate3d(-125%, 0, 0) scaleX(0.78);
@@ -1750,10 +1774,10 @@
 	margin: 0 !important;
 }
 
-/* Composer shadow on the dark background color so toasts
-   do not merge into card-colored surfaces behind them. */
+/* Dark toasts share the chatbox surface color (.chat-composer-surface
+   uses var(--card) in dark); the composer shadow lifts them off the page. */
 .dark [data-sonner-toast][data-styled='true'] {
-	background-color: var(--background) !important;
+	background-color: var(--card) !important;
 	box-shadow: 0 2px 8px -2px rgba(0, 0, 0, 0.16) !important;
 }
 
@@ -2099,6 +2123,11 @@
 		animation-iteration-count: infinite !important;
 	}
 
+	.loading-bar-slide {
+		animation-duration: 1.3s !important;
+		animation-iteration-count: infinite !important;
+	}
+
 	/* Keep the plus/x morph animating under reduced motion (small rotation,
 	   like the spinners above). */
 	.unsloth-composer-plus svg {

studio/install_llama_prebuilt.py

@@ -460,8 +460,14 @@ def is_busy_lock_error(exc: BaseException) -> bool:
     return False
 
 
+# Status logs default to stderr so resolver modes keep stdout machine-readable
+# (setup.sh json.load()s the whole stdout). main() flips this for the install
+# path, where PowerShell otherwise renders stderr as NativeCommandError noise.
+_LOG_TO_STDOUT = False
+
+
 def log(message: str) -> None:
-    print(f"[llama-prebuilt] {message}", file = sys.stderr)
+    print(f"[llama-prebuilt] {message}", file = sys.stdout if _LOG_TO_STDOUT else sys.stderr)
 
 
 def log_lines(lines: Iterable[str]) -> None:
@@ -856,6 +862,16 @@ def format_byte_count(num_bytes: float) -> str:
     return f"{num_bytes:.1f} B"
 
 
+def _progress_percent_step() -> int:
+    """Non-tty milestone granularity. The in-app updater sets
+    UNSLOTH_PROGRESS_PERCENT_STEP=5 to stream finer progress lines."""
+    try:
+        step = int(os.environ.get("UNSLOTH_PROGRESS_PERCENT_STEP", "25"))
+    except ValueError:
+        return 25
+    return min(max(step, 1), 50)
+
+
 class DownloadProgress:
     def __init__(self, label: str, total_bytes: int | None) -> None:
         self.label = label
@@ -868,6 +884,7 @@ class DownloadProgress:
         )
         self.is_tty = term_ok and self.stream.isatty()
         self.completed = False
+        self.milestone_step = _progress_percent_step()
         self.last_milestone_percent = -1
         self.last_milestone_bytes = 0
         self.has_rendered_tty_progress = False
@@ -918,7 +935,8 @@ class DownloadProgress:
         should_emit = False
         if self.total_bytes is not None:
             percent = int((downloaded_bytes * 100) / max(self.total_bytes, 1))
-            milestone_percent = min((percent // 25) * 25, 100)
+            step = self.milestone_step
+            milestone_percent = min((percent // step) * step, 100)
             if milestone_percent > self.last_milestone_percent and milestone_percent < 100:
                 self.last_milestone_percent = milestone_percent
                 should_emit = True
@@ -3916,6 +3934,12 @@ def _is_trusted_github_release_url(url: str, expected_repo: str) -> bool:
     return False
 
 
+# (gfx_target, asset_name) pairs already logged. resolve_lemonade_rocm_choice()
+# runs twice per install (direct planner + resolve_upstream_asset_choice), so
+# this stops its selection banner and hash-manifest NOTE printing twice.
+_lemonade_selection_logged: "set[tuple[str, str]]" = set()
+
+
 @functools.lru_cache(maxsize = 8)
 def _fetch_lemonade_release_cached(api_url: str, llama_tag: str) -> "dict | None":
     """Cached wrapper around fetch_json for lemonade release lookups.
@@ -4022,18 +4046,22 @@ def resolve_lemonade_rocm_choice(
     # Note: lemonade tags Linux assets with "ubuntu" but the binary is a
     # generic glibc build that runs on any distro (Arch, Fedora, ...), so
     # this attempt is selected for all Linux ROCm hosts, not just Ubuntu.
-    log(
-        f"AMD GPU {host.rocm_gfx_target!r} ({gfx_family}) -- "
-        f"trying lemonade-sdk ROCm prebuilt {asset_name} "
-        f"(works on any glibc Linux, not just Ubuntu)"
-    )
-    log(
-        f"NOTE: lemonade-sdk/llamacpp-rocm releases are not covered by the "
-        f"Unsloth approved-hash manifest; download integrity relies on "
-        f"functional validation (llama-bench / llama-server smoke tests) "
-        f"after extraction. Set UNSLOTH_DISABLE_LEMONADE_ROCM=1 to skip "
-        f"lemonade and fall back to the upstream HIP build path."
-    )
+    # Log once per (gfx_target, asset); see _lemonade_selection_logged.
+    log_key = (host.rocm_gfx_target, asset_name)
+    if log_key not in _lemonade_selection_logged:
+        _lemonade_selection_logged.add(log_key)
+        log(
+            f"AMD GPU {host.rocm_gfx_target!r} ({gfx_family}) -- "
+            f"trying lemonade-sdk ROCm prebuilt {asset_name} "
+            f"(works on any glibc Linux, not just Ubuntu)"
+        )
+        log(
+            f"NOTE: lemonade-sdk/llamacpp-rocm releases are not covered by the "
+            f"Unsloth approved-hash manifest; download integrity relies on "
+            f"functional validation (llama-bench / llama-server smoke tests) "
+            f"after extraction. Set UNSLOTH_DISABLE_LEMONADE_ROCM=1 to skip "
+            f"lemonade and fall back to the upstream HIP build path."
+        )
     return AssetChoice(
         repo = LEMONADE_ROCM_REPO,
         tag = release_tag,
@@ -7012,6 +7040,9 @@ def main() -> int:
         raise SystemExit(
             "install_llama_prebuilt.py: --install-dir is required unless --resolve-llama-tag, --resolve-install-tag, or --resolve-source-build is used"
         )
+    # Install path only: route status logs to stdout (see _LOG_TO_STDOUT note).
+    global _LOG_TO_STDOUT
+    _LOG_TO_STDOUT = True
     install_prebuilt(
         install_dir = Path(args.install_dir).expanduser().resolve(),
         llama_tag = args.llama_tag,

tests/utils/test_rope_scaling_drift.py

@@ -300,7 +300,14 @@ def test_object_style_rope_scaling_on_config_delegates_correctly():
     expected = _reference_inv_freq(dict_config, "linear")
 
     object_config = _make_config({"rope_type": "linear", "factor": 4.0})
-    object_config.rope_scaling = FakeLinearRopeScalingConfig()
+    try:
+        object_config.rope_scaling = FakeLinearRopeScalingConfig()
+    except Exception:
+        pytest.skip(
+            "transformers strict-validates rope_scaling to dict/RopeParameters/None, "
+            "so object-style config.rope_scaling (and the delegation retry it "
+            "exercises) is unreachable on this version."
+        )
     inv_freq, attention_scaling = _compute_config_rope_inv_freq(
         object_config, object_config.rope_scaling
     )

unsloth/models/_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2026.6.2"
+__version__ = "2026.6.3"
 
 __all__ = [
     "SUPPORTS_BFLOAT16",

unsloth/models/llama.py

@@ -1673,12 +1673,26 @@ def _llama3_inv_freq_from_config(
     return torch.where(is_medium, smoothed, scaled)
 
 
+def _vanilla_inv_freq_from_config(config, device = "cpu"):
+    """Unscaled RoPE inv_freq (rope_type 'default'/None), matching the constructor's fallback."""
+    base = _get_rope_theta(config, default = 10000.0)
+    dim = getattr(config, "head_dim", None)
+    if dim is None:
+        dim = int(config.hidden_size // config.num_attention_heads)
+    return 1.0 / (base ** (torch.arange(0, dim, 2, dtype = torch.int64, device = device).float() / dim))
+
+
 def _compute_config_rope_inv_freq(config, rope_scaling):
     """(inv_freq, attention_scaling) per config.rope_scaling via transformers'
     ROPE_INIT_FUNCTIONS, with an inline llama3 fallback; (None, 1.0) on failure."""
     original_rope_scaling = rope_scaling
     rope_scaling = _rope_scaling_as_dict(rope_scaling)
     rope_type = rope_scaling.get("rope_type", None) or rope_scaling.get("type", None)
+    # "default"/unset means unscaled RoPE. transformers >=5 reports
+    # rope_type="default" for every plain config and dropped "default" from
+    # ROPE_INIT_FUNCTIONS, so compute it directly instead of warning per load.
+    if rope_type in (None, "default"):
+        return _vanilla_inv_freq_from_config(config).to(dtype = torch.float32, device = "cpu"), 1.0
     try:
         from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
 

unsloth_cli/commands/studio.py

@@ -570,6 +570,19 @@ def _load_model_via_http(
         raise RuntimeError(f"Model load failed (HTTP {exc.code}): {body}") from exc
 
 
+def _format_context_length_line(load_result: dict) -> Optional[str]:
+    value = load_result.get("context_length")
+    if isinstance(value, bool):
+        return None
+    try:
+        value_int = int(value)
+    except (TypeError, ValueError):
+        return None
+    if value_int <= 0:
+        return None
+    return f"  Context length: {value_int} tokens"
+
+
 # ── unsloth studio (server) ──────────────────────────────────────────
 
 
@@ -841,7 +854,10 @@ def run(
         None, "--gguf-variant", help = "GGUF quant variant (e.g. UD-Q4_K_XL)"
     ),
     max_seq_length: int = typer.Option(
-        0, "--max-seq-length", help = "Max sequence length (0 = model default)"
+        0,
+        "--max-seq-length",
+        "--context-length",
+        help = "Runtime context length in tokens (0 = model default for GGUF; 2048 for hub models)",
     ),
     load_in_4bit: bool = typer.Option(True, "--load-in-4bit/--no-load-in-4bit"),
     api_key_name: str = typer.Option(
@@ -1080,6 +1096,7 @@ def run(
 
     loaded_model = result.get("model", model)
     display_variant = f" ({gguf_variant})" if gguf_variant else ""
+    context_length_line = _format_context_length_line(result)
 
     # 6. Print banner.
     display_host = _resolve_external_ip() if host == "0.0.0.0" else host
@@ -1119,6 +1136,8 @@ def run(
         if _cf_url:
             typer.echo(f"  Secure link access via Cloudflare: {_cf_url}")
         typer.echo(f"  Model loaded: {loaded_model}{display_variant}")
+        if context_length_line:
+            typer.echo(context_length_line)
         typer.echo(f"  API Key:      {api_key}")
         typer.echo("")
         typer.echo("  OpenAI / Anthropic SDK base URL:")
@@ -1153,6 +1172,8 @@ def run(
         typer.echo(f"URL:     {base_url}")
         if _cf_url:
             typer.echo(f"Secure link access via Cloudflare: {_cf_url}")
+        if context_length_line:
+            typer.echo(context_length_line.strip())
         typer.echo(f"API Key: {api_key}")
         typer.secho(_tool_notice, fg = _tool_notice_fg, bold = True)
 

unsloth_cli/tests/test_studio_run_parallel_flag.py

@@ -51,6 +51,18 @@ def test_parallel_option_is_registered():
         assert required in flags, f"flag {required!r} missing from --parallel option"
 
 
+def test_context_length_alias_is_registered():
+    """`--context-length` is an operator-facing alias for --max-seq-length."""
+    studio_mod = _load_run_command()
+    import inspect
+
+    sig = inspect.signature(studio_mod.run)
+    opt = sig.parameters["max_seq_length"].default
+    flags = set(getattr(opt, "param_decls", None) or [])
+    assert "--max-seq-length" in flags
+    assert "--context-length" in flags
+
+
 def test_parallel_default_is_four():
     """Default must stay at 4 so plain `unsloth studio run` is unchanged."""
     studio_mod = _load_run_command()
@@ -249,6 +261,15 @@ def test_reexec_np_is_first_class_alias(monkeypatch):
     assert _value_after(argv, "--port") == "8888", argv
 
 
+def test_reexec_forwards_context_length_alias(monkeypatch):
+    """Alias should normalize to the existing child --max-seq-length flag."""
+    result, captured = _invoke_run(monkeypatch, _BASE + ["--context-length", "8192"])
+    assert len(captured) == 1, result.output
+    argv = captured[0]["argv"]
+    assert _value_after(argv, "--max-seq-length") == "8192", argv
+    assert "--context-length" not in argv, argv
+
+
 def test_reexec_mixed_parallel_with_passthrough(monkeypatch):
     """--parallel + llama-server pass-through flags must all reach the child."""
     result, captured = _invoke_run(
@@ -262,6 +283,22 @@ def test_reexec_mixed_parallel_with_passthrough(monkeypatch):
     assert _value_after(argv, "--temp") == "0.7", argv
 
 
+def test_context_length_banner_line_formats_ints():
+    studio_mod = _load_run_command()
+    assert studio_mod._format_context_length_line({"context_length": 4096}) == (
+        "  Context length: 4096 tokens"
+    )
+    assert studio_mod._format_context_length_line({"context_length": "8192"}) == (
+        "  Context length: 8192 tokens"
+    )
+
+
+@pytest.mark.parametrize("value", [None, 0, -1, True, ""])
+def test_context_length_banner_line_omits_unknown_values(value):
+    studio_mod = _load_run_command()
+    assert studio_mod._format_context_length_line({"context_length": value}) is None
+
+
 @pytest.mark.parametrize(
     "user_flag,expected_in_child",
     [