diff --git a/install.sh b/install.sh
index d12abe298f..49bb7a7b89 100755
--- a/install.sh
+++ b/install.sh
@@ -1290,6 +1290,14 @@ if [ "$_NO_TORCH_FLAG" = true ] || [ "$MAC_INTEL" = true ]; then
     SKIP_TORCH=true
 fi
 
+# Apple Silicon: override mlx-vlm / mlx-lm's transformers pin (see overrides file).
+if [ "$OS" = "macos" ] && [ "$_ARCH" = "arm64" ]; then
+    _OVERRIDES_FILE="$(cd "$(dirname "$0" 2>/dev/null || echo ".")" && pwd)/studio/backend/requirements/single-env/overrides-darwin-arm64.txt"
+    if [ -f "$_OVERRIDES_FILE" ]; then
+        export UV_OVERRIDE="$_OVERRIDES_FILE"
+    fi
+fi
+
 _TAURI_INITIAL_GPU_BRANCH="unknown"
 if [ "$SKIP_TORCH" = true ]; then
     _TAURI_INITIAL_GPU_BRANCH="no_torch"
@@ -2108,12 +2116,6 @@ else
     fi
 fi
 
-# ── Install mlx-vlm on Apple Silicon (optional, for VLM training) ──
-if [ "$OS" = "macos" ] && [ "$_ARCH" = "arm64" ]; then
-    substep "installing mlx-vlm (VLM training support)..."
-    run_install_cmd "install mlx-vlm" uv pip install --python "$_VENV_PY" mlx-vlm
-fi
-
 # ── Run studio setup ──
 tauri_log "STEP" "Running Studio setup"
 # When --local, use the repo's own setup.sh directly.
diff --git a/studio/backend/requirements/single-env/overrides-darwin-arm64.txt b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
new file mode 100644
index 0000000000..2cd03d8b78
--- /dev/null
+++ b/studio/backend/requirements/single-env/overrides-darwin-arm64.txt
@@ -0,0 +1,5 @@
+# mlx-vlm / mlx-lm declare transformers>=5.x which conflicts with the
+# main venv's constraints.txt pin transformers==4.57.6 and forces uv to
+# backtrack unsloth. Relax to match the pin -- per-model 5.x routing
+# happens at runtime via the side-car venvs.
+transformers>=4.57.6
diff --git a/studio/backend/utils/hardware/hardware.py b/studio/backend/utils/hardware/hardware.py
index 3764e38272..ede37e2953 100644
--- a/studio/backend/utils/hardware/hardware.py
+++ b/studio/backend/utils/hardware/hardware.py
@@ -144,7 +144,10 @@ def detect_hardware() -> DeviceType:
     if is_apple_silicon() and _has_mlx():
         DEVICE = DeviceType.MLX
         CHAT_ONLY = False
-        chip = platform.processor() or platform.machine()
+        # platform.processor() runs `uname -p` which returns "i386" on most
+        # universal2 / Rosetta-shaped Python builds even on native arm64.
+        # platform.machine() is "arm64" once is_apple_silicon() has gated us.
+        chip = platform.machine() or "arm64"
         print(f"Hardware detected: MLX — Apple Silicon ({chip})")
         return DEVICE
 
@@ -279,13 +282,11 @@ def get_gpu_memory_info() -> Dict[str, Any]:
 
             try:
                 info = mx.device_info()
-                gpu_name = (
-                    info.get("device_name")
-                    or platform.processor()
-                    or platform.machine()
-                )
+                # See detect_hardware(): platform.processor() can return "i386"
+                # on native arm64 Python builds, so prefer machine() as fallback.
+                gpu_name = info.get("device_name") or platform.machine() or "arm64"
             except Exception:
-                gpu_name = platform.processor() or platform.machine()
+                gpu_name = platform.machine() or "arm64"
 
             return {
                 "available": True,
diff --git a/studio/install_python_stack.py b/studio/install_python_stack.py
index 9166d35ce3..ca7fe3f004 100644
--- a/studio/install_python_stack.py
+++ b/studio/install_python_stack.py
@@ -37,6 +37,7 @@ from backend.utils.wheel_utils import (
 IS_WINDOWS = sys.platform == "win32"
 IS_MACOS = sys.platform == "darwin"
 IS_MAC_INTEL = IS_MACOS and platform.machine() == "x86_64"
+IS_MAC_ARM = IS_MACOS and platform.machine() == "arm64"
 
 # ── ROCm / AMD GPU support ─────────────────────────────────────────────────────
 # Mapping from detected ROCm (major, minor) to the best PyTorch wheel tag on
@@ -423,6 +424,7 @@ def _infer_no_torch() -> bool:
 
 NO_TORCH = _infer_no_torch()
 
+
 # -- Verbosity control ----------------------------------------------------------
 # By default the installer shows a minimal progress bar (one line, in-place).
 # Set UNSLOTH_VERBOSE=1 in the environment to restore full per-step output:
@@ -448,6 +450,11 @@ LOCAL_DD_GITHUB_PLUGIN = (
     SCRIPT_DIR / "backend" / "plugins" / "data-designer-github-repo-seed"
 )
 
+# Apple Silicon: override mlx-vlm/mlx-lm's transformers pin (see overrides file).
+_MLX_OVERRIDES = SINGLE_ENV / "overrides-darwin-arm64.txt"
+if IS_MAC_ARM and _MLX_OVERRIDES.is_file():
+    os.environ.setdefault("UV_OVERRIDE", str(_MLX_OVERRIDES))
+
 # -- Unicode-safe printing ---------------------------------------------
 # On Windows the default console encoding can be a legacy code page
 # (e.g. CP1252) that cannot represent Unicode glyphs such as ✅ or ❌.
@@ -960,6 +967,20 @@ def install_python_stack() -> int:
                 [sys.executable, "-m", "pip", "install", "--upgrade", "pip"],
             )
 
+    # macOS arm64: install MLX stack at latest (UV_OVERRIDE relaxes the
+    # mlx-vlm / mlx-lm transformers pin -- set at module load).
+    if IS_MAC_ARM and not skip_base:
+        _progress("MLX stack (Apple Silicon)")
+        pip_install(
+            "Installing MLX stack (mlx + mlx-lm + mlx-vlm)",
+            "--no-cache-dir",
+            "--upgrade",
+            "mlx",
+            "mlx-metal",
+            "mlx-lm",
+            "mlx-vlm",
+        )
+
     # 3. Core packages: unsloth-zoo + unsloth (or custom package name)
     if skip_base:
         pass