From 646460b25a847308ab6bf0f2a2d2aeea5228960b Mon Sep 17 00:00:00 2001
From: XXXXRT666 <157766680+XXXXRT666@users.noreply.github.com>
Date: Wed, 19 Aug 2026 01:42:23 +0800
Subject: [PATCH 1/6] Update nanobind to 3.0.0-dev1

---
 CMakeLists.txt                                | 11 ++++--
 .../apply-nanobind-v3.0.0-dev1-gil.cmake      | 34 +++++++++++++++++++
 cmake/patches/nanobind-v3.0.0-dev1-gil.patch  | 13 +++++++
 docs/src/dev/extensions.rst                   |  2 +-
 examples/extensions/CMakeLists.txt            |  2 +-
 examples/extensions/pyproject.toml            |  2 +-
 examples/extensions/requirements.txt          |  2 +-
 python/src/small_vector.h                     |  2 +-
 python/src/transforms.cpp                     |  8 ++---
 9 files changed, 65 insertions(+), 11 deletions(-)
 create mode 100644 cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake
 create mode 100644 cmake/patches/nanobind-v3.0.0-dev1-gil.patch

diff --git a/CMakeLists.txt b/CMakeLists.txt
index feb7ce5ebb..3193dad4e5 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -395,9 +395,16 @@ if(MLX_BUILD_PYTHON_BINDINGS)
   FetchContent_Declare(
     nanobind
     GIT_REPOSITORY https://github.com/wjakob/nanobind.git
-    GIT_TAG v2.15.0
+    GIT_TAG v3.0.0-dev1
     GIT_SHALLOW TRUE
-    EXCLUDE_FROM_ALL)
+    EXCLUDE_FROM_ALL
+    # v3.0.0-dev1 mistakes a non-null CPython < 3.13 thread state for GIL
+    # ownership. Drop this patch once the fix is included in the pinned nanobind
+    # tag.
+    PATCH_COMMAND
+      ${CMAKE_COMMAND} "-DNANOBIND_SOURCE_DIR=<SOURCE_DIR>" -P
+      ${CMAKE_CURRENT_SOURCE_DIR}/cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake
+  )
   FetchContent_MakeAvailable(nanobind)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/python/src)
 endif()
diff --git a/cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake b/cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake
new file mode 100644
index 0000000000..5b6adcdd1a
--- /dev/null
+++ b/cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake
@@ -0,0 +1,34 @@
+if(NOT DEFINED NANOBIND_SOURCE_DIR)
+  message(FATAL_ERROR "NANOBIND_SOURCE_DIR is required")
+endif()
+
+set(patch_file "${CMAKE_CURRENT_LIST_DIR}/nanobind-v3.0.0-dev1-gil.patch")
+
+execute_process(
+  COMMAND git apply --check "${patch_file}"
+  WORKING_DIRECTORY "${NANOBIND_SOURCE_DIR}"
+  RESULT_VARIABLE can_apply
+  OUTPUT_QUIET ERROR_QUIET)
+
+if(can_apply EQUAL 0)
+  execute_process(
+    COMMAND git apply "${patch_file}"
+    WORKING_DIRECTORY "${NANOBIND_SOURCE_DIR}"
+    RESULT_VARIABLE apply_result
+    ERROR_VARIABLE apply_error)
+  if(NOT apply_result EQUAL 0)
+    message(FATAL_ERROR "Failed to apply nanobind GIL patch: ${apply_error}")
+  endif()
+  return()
+endif()
+
+execute_process(
+  COMMAND git apply --reverse --check "${patch_file}"
+  WORKING_DIRECTORY "${NANOBIND_SOURCE_DIR}"
+  RESULT_VARIABLE already_applied
+  OUTPUT_QUIET ERROR_QUIET)
+
+if(NOT already_applied EQUAL 0)
+  message(
+    FATAL_ERROR "nanobind source is neither patchable nor already patched")
+endif()
diff --git a/cmake/patches/nanobind-v3.0.0-dev1-gil.patch b/cmake/patches/nanobind-v3.0.0-dev1-gil.patch
new file mode 100644
index 0000000000..49aae8020b
--- /dev/null
+++ b/cmake/patches/nanobind-v3.0.0-dev1-gil.patch
@@ -0,0 +1,13 @@
+diff --git a/src/nb_internals.h b/src/nb_internals.h
+index 43881a4..e3c9326 100644
+--- a/src/nb_internals.h
++++ b/src/nb_internals.h
+@@ -937,7 +937,7 @@ NB_INLINE bool tstate_attached() noexcept {
+ #if defined(Py_LIMITED_API) || defined(PYPY_VERSION)
+     return false;
+ #elif PY_VERSION_HEX < 0x030D0000
+-    return _PyThreadState_UncheckedGet() != nullptr;
++    return PyGILState_Check() != 0;
+ #else
+     return PyThreadState_GetUnchecked() != nullptr;
+ #endif
diff --git a/docs/src/dev/extensions.rst b/docs/src/dev/extensions.rst
index 89c8991e56..5f5fab6429 100644
--- a/docs/src/dev/extensions.rst
+++ b/docs/src/dev/extensions.rst
@@ -711,7 +711,7 @@ build utilities defined in :mod:`mlx.extension`:
             package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
             extras_require={"dev":[]},
             zip_safe=False,
-            python_requires=">=3.8",
+            python_requires=">=3.10",
         )
 
 .. note::
diff --git a/examples/extensions/CMakeLists.txt b/examples/extensions/CMakeLists.txt
index eedc3a7dab..50422261e3 100644
--- a/examples/extensions/CMakeLists.txt
+++ b/examples/extensions/CMakeLists.txt
@@ -11,7 +11,7 @@ option(BUILD_SHARED_LIBS "Build extensions as a shared library" ON)
 
 # ----------------------------- Dependencies -----------------------------
 find_package(
-  Python 3.8
+  Python 3.10
   COMPONENTS Interpreter Development.Module
   REQUIRED)
 execute_process(
diff --git a/examples/extensions/pyproject.toml b/examples/extensions/pyproject.toml
index 560a58bc28..3d8f06320f 100644
--- a/examples/extensions/pyproject.toml
+++ b/examples/extensions/pyproject.toml
@@ -3,6 +3,6 @@ requires = [
   "setuptools>=42",
   "cmake>=3.25",
   "mlx>=0.18.0",
-  "nanobind==2.15.0",
+  "nanobind==3.0.0.dev1",
 ]
 build-backend = "setuptools.build_meta"
diff --git a/examples/extensions/requirements.txt b/examples/extensions/requirements.txt
index 917d125eea..e130096c26 100644
--- a/examples/extensions/requirements.txt
+++ b/examples/extensions/requirements.txt
@@ -1,4 +1,4 @@
 setuptools>=42
 cmake>=3.25
 mlx>=0.31.2
-nanobind==2.15.0
+nanobind==3.0.0.dev1
diff --git a/python/src/small_vector.h b/python/src/small_vector.h
index 0b2e6bb2ee..1a07abe1b2 100644
--- a/python/src/small_vector.h
+++ b/python/src/small_vector.h
@@ -31,7 +31,7 @@ struct type_caster<::mlx::core::SmallVector<Type, Size, Alloc>> {
 
   // Not noexcept: on overflow of a narrow integer element we raise
   // OverflowError so nanobind surfaces a clean error to the user.
-  bool from_python(handle src, uint8_t flags, cleanup_list* cleanup) {
+  bool from_python(handle src, uint32_t flags, cleanup_list* cleanup) {
     size_t size;
     PyObject* temp;
 
diff --git a/python/src/transforms.cpp b/python/src/transforms.cpp
index 1ec20a1375..2d9e2d3467 100644
--- a/python/src/transforms.cpp
+++ b/python/src/transforms.cpp
@@ -809,8 +809,8 @@ class PyCustomFunction {
       }
       int array_index = 0;
       int tangent_index = 0;
-      auto new_tangents =
-          nb::cast<nb::tuple>(tree_map(args, [&](nb::handle element) {
+      auto new_tangents = nb::cast<nb::tuple>(
+          tree_map(args, [&](nb::handle element) -> nb::object {
             if (nb::isinstance<mx::array>(element) &&
                 have_tangents[array_index++]) {
               return nb::cast(tangents[tangent_index++]);
@@ -856,8 +856,8 @@ class PyCustomFunction {
       }
 
       int arr_index = 0;
-      auto new_axes =
-          nb::cast<nb::tuple>(tree_map(args, [&](nb::handle element) {
+      auto new_axes = nb::cast<nb::tuple>(
+          tree_map(args, [&](nb::handle element) -> nb::object {
             int axis = axes[arr_index++];
             if (nb::isinstance<mx::array>(element) && axis >= 0) {
               return nb::cast(axis);

From e2c0781c8809535670acc8fa05c12b8feeb310a2 Mon Sep 17 00:00:00 2001
From: XXXXRT666 <157766680+XXXXRT666@users.noreply.github.com>
Date: Wed, 19 Aug 2026 21:13:07 +0800
Subject: [PATCH 2/6] Enable nanobind split mode and free-threaded wheels

---
 .github/actions/test-wheel/action.yml         | 15 ++++++--
 .github/workflows/release.yml                 | 37 +++++++++++++------
 CMakeLists.txt                                | 29 ++++++++++-----
 MANIFEST.in                                   |  1 +
 _build_backend/backend.py                     | 23 ++++++++++++
 .../apply-nanobind-v3.0.0-dev1-gil.cmake      | 34 -----------------
 cmake/patches/nanobind-v3.0.0-dev1-gil.patch  | 13 -------
 docs/src/dev/extensions.rst                   |  7 +++-
 examples/extensions/CMakeLists.txt            |  6 +--
 examples/extensions/pyproject.toml            |  5 ++-
 examples/extensions/requirements.txt          |  5 ++-
 examples/extensions/setup.py                  |  2 +
 pyproject.toml                                |  5 ++-
 python/mlx/extension.py                       |  2 +-
 python/src/CMakeLists.txt                     | 12 ++++--
 python/src/buffer.h                           | 22 +++++++++++
 python/src/load.cpp                           |  6 +++
 python/src/mlx_func.cpp                       | 16 ++++++--
 python/src/small_vector.h                     |  2 +-
 setup.py                                      | 13 ++++++-
 20 files changed, 163 insertions(+), 92 deletions(-)
 create mode 100644 _build_backend/backend.py
 delete mode 100644 cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake
 delete mode 100644 cmake/patches/nanobind-v3.0.0-dev1-gil.patch

diff --git a/.github/actions/test-wheel/action.yml b/.github/actions/test-wheel/action.yml
index b745a4d2f9..35e8c3c7d5 100644
--- a/.github/actions/test-wheel/action.yml
+++ b/.github/actions/test-wheel/action.yml
@@ -10,15 +10,24 @@ inputs:
 runs:
   using: 'composite'
   steps:
-    - name: Get Python version
+    - name: Get frontend ABI
       id: python
       shell: bash
-      run: python -c "import sys; print(f'version={sys.version_info.major}.{sys.version_info.minor}')" >> $GITHUB_OUTPUT
+      run: |
+        python - <<'PY' >> "$GITHUB_OUTPUT"
+        import sys
+        import sysconfig
+
+        if sysconfig.get_config_var("Py_GIL_DISABLED"):
+            print(f"abi=py{sys.version_info.major}.{sys.version_info.minor}t")
+        else:
+            print("abi=abi3")
+        PY
 
     - name: Download frontend packages
       uses: actions/download-artifact@v8
       with:
-        pattern: frontend-${{ runner.os }}-${{ runner.arch }}-py${{ steps.python.outputs.version }}
+        pattern: frontend-${{ runner.os }}-${{ runner.arch }}-${{ steps.python.outputs.abi }}
         path: wheelhouse
 
     - name: Download backend packages
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index b53f003ff4..2b2aa91380 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -58,12 +58,18 @@ jobs:
       matrix:
         os: ['Linux', 'Windows']
         arch: ['x86_64', 'aarch64']
-        python-version: &pyver ['3.10', '3.11', '3.12', '3.13', '3.13t', '3.14', '3.14t']
-        # There is no cp310 binary for Windows on arm.
+        # One Stable ABI frontend wheel covers all supported regular CPython
+        # versions on a platform. Python 3.13t and 3.14t do not have a
+        # free-threaded Stable ABI and require version-specific linked wheels.
+        python-version: &pyver ['3.10', '3.13t', '3.14t']
         exclude:
           - os: 'Windows'
             arch: 'aarch64'
             python-version: '3.10'
+        include:
+          - os: 'Windows'
+            arch: 'aarch64'
+            python-version: '3.11'
     runs-on: |-
       ${{ case(matrix.os == 'Windows',
                case(matrix.arch == 'aarch64', 'windows-11-arm',
@@ -89,7 +95,7 @@ jobs:
       - run: unzip -l wheelhouse/*.whl
       - uses: actions/upload-artifact@v7
         with:
-          name: frontend-${{ runner.os }}-${{ runner.arch }}-py${{ matrix.python-version }}
+          name: frontend-${{ runner.os }}-${{ runner.arch }}-${{ contains(matrix.python-version, 't') && format('py{0}', matrix.python-version) || 'abi3' }}
           path: wheelhouse/mlx-*.whl
           if-no-files-found: error
 
@@ -175,7 +181,7 @@ jobs:
       - name: Upload frontend packages
         uses: actions/upload-artifact@v7
         with:
-          name: frontend-${{ runner.os }}-${{ runner.arch }}-py${{ matrix.python-version }}
+          name: frontend-${{ runner.os }}-${{ runner.arch }}-${{ contains(matrix.python-version, 't') && format('py{0}', matrix.python-version) || 'abi3' }}
           path: wheelhouse/mlx-*.whl
           if-no-files-found: error
       - name: Upload backend packages
@@ -187,30 +193,36 @@ jobs:
           if-no-files-found: error
 
   test_wheel:
-    name: Test (${{ matrix.os }}, ${{ matrix.toolkit }}, ${{ matrix.arch }})
+    name: Test (${{ matrix.os }}, python-${{ matrix.python-version }}, ${{ matrix.toolkit }}, ${{ matrix.arch }})
     if: github.repository == 'ml-explore/mlx'
     needs: [build_frontend, build_backend, build_mac_wheels]
     strategy:
       matrix:
         os: ['Linux', 'Windows']
-        arch: ['aarch64']
+        arch: ['x86_64', 'aarch64']
         toolkit: ['cpu']
+        python-version: ['3.12', '3.13t', '3.14t']
         include:
-          - os: 'Linux'
-            arch: 'x86_64'
-            toolkit: 'cpu'
           - os: 'Linux'
             arch: 'x86_64'
             toolkit: 'cuda-12.9'
+            python-version: '3.12'
           - os: 'Linux'
             arch: 'x86_64'
             toolkit: 'cuda-13.0'
-          - os: 'Windows'
-            arch: 'x86_64'
-            toolkit: 'cpu'
+            python-version: '3.12'
           - os: 'macOS'
             arch: 'aarch64'
             toolkit: 'metal'
+            python-version: '3.12'
+          - os: 'macOS'
+            arch: 'aarch64'
+            toolkit: 'metal'
+            python-version: '3.13t'
+          - os: 'macOS'
+            arch: 'aarch64'
+            toolkit: 'metal'
+            python-version: '3.14t'
     runs-on: |-
       ${{ case(matrix.os == 'Windows', case(matrix.arch == 'aarch64', 'windows-11-arm',
                                             'windows-2022'),
@@ -224,6 +236,7 @@ jobs:
       - uses: ./.github/actions/setup
         with:
           toolkit: ${{ matrix.toolkit }}
+          python-version: ${{ matrix.python-version }}
           use-ccache: false
       - uses: ./.github/actions/test-wheel
         with:
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3193dad4e5..3da61b3c47 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -1,4 +1,4 @@
-cmake_minimum_required(VERSION 3.25)
+cmake_minimum_required(VERSION 3.26)
 
 if(NOT MLX_VERSION)
   file(STRINGS "mlx/version.h" _mlx_h_version REGEX "^#define MLX_VERSION_.*$")
@@ -392,19 +392,28 @@ if(MLX_BUILD_PYTHON_BINDINGS)
     Python 3.10
     COMPONENTS Interpreter Development.Module
     REQUIRED)
+  execute_process(
+    COMMAND
+      "${Python_EXECUTABLE}" -c
+      "import sysconfig; print('ON' if sysconfig.get_config_var('Py_GIL_DISABLED') else 'OFF')"
+    RESULT_VARIABLE _mlx_python_gil_result
+    OUTPUT_VARIABLE MLX_PYTHON_FREE_THREADED
+    OUTPUT_STRIP_TRAILING_WHITESPACE)
+  if(NOT _mlx_python_gil_result EQUAL 0)
+    message(FATAL_ERROR "Failed to determine whether Python is free-threaded.")
+  endif()
+  if(NOT MLX_PYTHON_FREE_THREADED)
+    find_package(
+      Python 3.10
+      COMPONENTS Development.SABIModule
+      REQUIRED)
+  endif()
   FetchContent_Declare(
     nanobind
     GIT_REPOSITORY https://github.com/wjakob/nanobind.git
-    GIT_TAG v3.0.0-dev1
+    GIT_TAG v3.0.0-dev2
     GIT_SHALLOW TRUE
-    EXCLUDE_FROM_ALL
-    # v3.0.0-dev1 mistakes a non-null CPython < 3.13 thread state for GIL
-    # ownership. Drop this patch once the fix is included in the pinned nanobind
-    # tag.
-    PATCH_COMMAND
-      ${CMAKE_COMMAND} "-DNANOBIND_SOURCE_DIR=<SOURCE_DIR>" -P
-      ${CMAKE_CURRENT_SOURCE_DIR}/cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake
-  )
+    EXCLUDE_FROM_ALL)
   FetchContent_MakeAvailable(nanobind)
   add_subdirectory(${CMAKE_CURRENT_LIST_DIR}/python/src)
 endif()
diff --git a/MANIFEST.in b/MANIFEST.in
index 632fae78c6..6a3cf1af8a 100644
--- a/MANIFEST.in
+++ b/MANIFEST.in
@@ -1,5 +1,6 @@
 include CMakeLists.txt
 include mlx.pc.in
+include _build_backend/backend.py
 recursive-include mlx *
 include cmake/*
 include python/src/*
diff --git a/_build_backend/backend.py b/_build_backend/backend.py
new file mode 100644
index 0000000000..d4113225d1
--- /dev/null
+++ b/_build_backend/backend.py
@@ -0,0 +1,23 @@
+import os
+import sysconfig
+
+from setuptools import build_meta as _setuptools_build_meta
+from setuptools.build_meta import *  # type: ignore # noqa: F403
+
+
+def _with_nanobind_backend(requires):
+    build_backend = int(os.environ.get("MLX_BUILD_BACKEND_PACKAGE", 0))
+    free_threaded = bool(sysconfig.get_config_var("Py_GIL_DISABLED"))
+    if not build_backend and not free_threaded:
+        requires.append("nanobind-backend>=1.0.0.dev2")
+    return requires
+
+
+def get_requires_for_build_wheel(config_settings=None):
+    requires = _setuptools_build_meta.get_requires_for_build_wheel(config_settings)
+    return _with_nanobind_backend(requires)
+
+
+def get_requires_for_build_editable(config_settings=None):
+    requires = _setuptools_build_meta.get_requires_for_build_editable(config_settings)
+    return _with_nanobind_backend(requires)
diff --git a/cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake b/cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake
deleted file mode 100644
index 5b6adcdd1a..0000000000
--- a/cmake/patches/apply-nanobind-v3.0.0-dev1-gil.cmake
+++ /dev/null
@@ -1,34 +0,0 @@
-if(NOT DEFINED NANOBIND_SOURCE_DIR)
-  message(FATAL_ERROR "NANOBIND_SOURCE_DIR is required")
-endif()
-
-set(patch_file "${CMAKE_CURRENT_LIST_DIR}/nanobind-v3.0.0-dev1-gil.patch")
-
-execute_process(
-  COMMAND git apply --check "${patch_file}"
-  WORKING_DIRECTORY "${NANOBIND_SOURCE_DIR}"
-  RESULT_VARIABLE can_apply
-  OUTPUT_QUIET ERROR_QUIET)
-
-if(can_apply EQUAL 0)
-  execute_process(
-    COMMAND git apply "${patch_file}"
-    WORKING_DIRECTORY "${NANOBIND_SOURCE_DIR}"
-    RESULT_VARIABLE apply_result
-    ERROR_VARIABLE apply_error)
-  if(NOT apply_result EQUAL 0)
-    message(FATAL_ERROR "Failed to apply nanobind GIL patch: ${apply_error}")
-  endif()
-  return()
-endif()
-
-execute_process(
-  COMMAND git apply --reverse --check "${patch_file}"
-  WORKING_DIRECTORY "${NANOBIND_SOURCE_DIR}"
-  RESULT_VARIABLE already_applied
-  OUTPUT_QUIET ERROR_QUIET)
-
-if(NOT already_applied EQUAL 0)
-  message(
-    FATAL_ERROR "nanobind source is neither patchable nor already patched")
-endif()
diff --git a/cmake/patches/nanobind-v3.0.0-dev1-gil.patch b/cmake/patches/nanobind-v3.0.0-dev1-gil.patch
deleted file mode 100644
index 49aae8020b..0000000000
--- a/cmake/patches/nanobind-v3.0.0-dev1-gil.patch
+++ /dev/null
@@ -1,13 +0,0 @@
-diff --git a/src/nb_internals.h b/src/nb_internals.h
-index 43881a4..e3c9326 100644
---- a/src/nb_internals.h
-+++ b/src/nb_internals.h
-@@ -937,7 +937,7 @@ NB_INLINE bool tstate_attached() noexcept {
- #if defined(Py_LIMITED_API) || defined(PYPY_VERSION)
-     return false;
- #elif PY_VERSION_HEX < 0x030D0000
--    return _PyThreadState_UncheckedGet() != nullptr;
-+    return PyGILState_Check() != 0;
- #else
-     return PyThreadState_GetUnchecked() != nullptr;
- #endif
diff --git a/docs/src/dev/extensions.rst b/docs/src/dev/extensions.rst
index 5f5fab6429..182408dcb1 100644
--- a/docs/src/dev/extensions.rst
+++ b/docs/src/dev/extensions.rst
@@ -679,7 +679,8 @@ Finally, we build the nanobind_ bindings
 
     nanobind_add_module(
       _ext
-      NB_STATIC STABLE_ABI LTO NOMINSIZE
+      BACKEND_MODULE nanobind_backend
+      LTO NOMINSIZE
       NB_DOMAIN mlx
       ${CMAKE_CURRENT_LIST_DIR}/bindings.cpp
     )
@@ -709,6 +710,8 @@ build utilities defined in :mod:`mlx.extension`:
             cmdclass={"build_ext": extension.CMakeBuild},
             packages=["mlx_sample_extensions"],
             package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
+            install_requires=["nanobind-backend>=1.0.0.dev2"],
+            options={"bdist_wheel": {"py_limited_api": "cp310"}},
             extras_require={"dev":[]},
             zip_safe=False,
             python_requires=">=3.10",
@@ -733,7 +736,7 @@ This results in the directory structure:
 | │   ├── __init__.py
 | │   ├── libmlx_ext.dylib # C++ extension library
 | │   ├── mlx_ext.metallib # Metal library
-| │   └── _ext.cpython-3x-darwin.so # Python Binding
+| │   └── _ext.abi3.so # Python Binding
 | ...
 
 When you try to install using the command ``python -m pip install .`` (in
diff --git a/examples/extensions/CMakeLists.txt b/examples/extensions/CMakeLists.txt
index 50422261e3..d7073b782d 100644
--- a/examples/extensions/CMakeLists.txt
+++ b/examples/extensions/CMakeLists.txt
@@ -12,7 +12,7 @@ option(BUILD_SHARED_LIBS "Build extensions as a shared library" ON)
 # ----------------------------- Dependencies -----------------------------
 find_package(
   Python 3.10
-  COMPONENTS Interpreter Development.Module
+  COMPONENTS Interpreter Development.Module Development.SABIModule
   REQUIRED)
 execute_process(
   COMMAND "${Python_EXECUTABLE}" -m nanobind --cmake_dir
@@ -64,8 +64,8 @@ endif()
 # ----------------------------- Python Bindings -----------------------------
 nanobind_add_module(
   _ext
-  NB_STATIC
-  STABLE_ABI
+  BACKEND_MODULE
+  nanobind_backend
   LTO
   NOMINSIZE
   NB_DOMAIN
diff --git a/examples/extensions/pyproject.toml b/examples/extensions/pyproject.toml
index 3d8f06320f..aa3c5ef2d5 100644
--- a/examples/extensions/pyproject.toml
+++ b/examples/extensions/pyproject.toml
@@ -1,8 +1,9 @@
 [build-system]
 requires = [
   "setuptools>=42",
-  "cmake>=3.25",
+  "cmake>=3.27",
   "mlx>=0.18.0",
-  "nanobind==3.0.0.dev1",
+  "nanobind==3.0.0.dev2",
+  "nanobind-backend>=1.0.0.dev2",
 ]
 build-backend = "setuptools.build_meta"
diff --git a/examples/extensions/requirements.txt b/examples/extensions/requirements.txt
index e130096c26..2a6b060a30 100644
--- a/examples/extensions/requirements.txt
+++ b/examples/extensions/requirements.txt
@@ -1,4 +1,5 @@
 setuptools>=42
-cmake>=3.25
+cmake>=3.27
 mlx>=0.31.2
-nanobind==3.0.0.dev1
+nanobind==3.0.0.dev2
+nanobind-backend>=1.0.0.dev2
diff --git a/examples/extensions/setup.py b/examples/extensions/setup.py
index be990a02b6..120d7050e8 100644
--- a/examples/extensions/setup.py
+++ b/examples/extensions/setup.py
@@ -13,6 +13,8 @@
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
+        install_requires=["nanobind-backend>=1.0.0.dev2"],
+        options={"bdist_wheel": {"py_limited_api": "cp310"}},
         zip_safe=False,
         python_requires=">=3.10",
     )
diff --git a/pyproject.toml b/pyproject.toml
index 45675ae856..f8f90c0e56 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,7 +1,8 @@
 [build-system]
 requires = [
   "setuptools>=80",
-  "cmake>=3.25",
+  "cmake>=3.26",
   "typing_extensions",
 ]
-build-backend = "setuptools.build_meta"
+build-backend = "backend"
+backend-path = ["_build_backend"]
diff --git a/python/mlx/extension.py b/python/mlx/extension.py
index d794fbe28f..c352ee4b42 100644
--- a/python/mlx/extension.py
+++ b/python/mlx/extension.py
@@ -17,7 +17,7 @@
 # A CMakeExtension needs a sourcedir instead of a file list.
 class CMakeExtension(Extension):
     def __init__(self, name: str, sourcedir: str = "") -> None:
-        super().__init__(name, sources=[])
+        super().__init__(name, sources=[], py_limited_api=True)
         self.sourcedir = os.fspath(Path(sourcedir).resolve())
 
 
diff --git a/python/src/CMakeLists.txt b/python/src/CMakeLists.txt
index 0798add410..141a35ec31 100644
--- a/python/src/CMakeLists.txt
+++ b/python/src/CMakeLists.txt
@@ -1,8 +1,14 @@
+# Python 3.13t and 3.14t have no free-threaded Stable ABI, so they must use
+# linked mode instead of the regular split-mode build.
+if(MLX_PYTHON_FREE_THREADED)
+  set(MLX_NANOBIND_MODE NB_STATIC FREE_THREADED)
+else()
+  set(MLX_NANOBIND_MODE BACKEND_MODULE nanobind_backend)
+endif()
+
 nanobind_add_module(
   core
-  NB_STATIC
-  STABLE_ABI
-  FREE_THREADED
+  ${MLX_NANOBIND_MODE}
   LTO
   NOMINSIZE
   NB_DOMAIN
diff --git a/python/src/buffer.h b/python/src/buffer.h
index 4b194b3d25..48dbf4ecfc 100644
--- a/python/src/buffer.h
+++ b/python/src/buffer.h
@@ -7,6 +7,28 @@
 #include "mlx/array.h"
 #include "mlx/utils.h"
 
+// Py_buffer became part of the Stable ABI in Python 3.11. CPython 3.10 uses
+// the same layout, which split-mode extensions targeting cp310 must declare.
+#if defined(Py_LIMITED_API) && Py_LIMITED_API + 0 < 0x030B0000
+typedef struct {
+  void* buf;
+  PyObject* obj;
+  Py_ssize_t len;
+  Py_ssize_t itemsize;
+  int readonly;
+  int ndim;
+  char* format;
+  Py_ssize_t* shape;
+  Py_ssize_t* strides;
+  Py_ssize_t* suboffsets;
+  void* internal;
+} Py_buffer;
+
+#define PyBUF_FORMAT 0x0004
+#define PyBUF_ND 0x0008
+#define PyBUF_STRIDES (0x0010 | PyBUF_ND)
+#endif
+
 // Only defined in >= Python 3.9
 // https://github.com/python/cpython/blob/f6cdc6b4a191b75027de342aa8b5d344fb31313e/Include/typeslots.h#L2-L3
 #ifndef Py_bf_getbuffer
diff --git a/python/src/load.cpp b/python/src/load.cpp
index 605049864e..39e8a1ab3a 100644
--- a/python/src/load.cpp
+++ b/python/src/load.cpp
@@ -15,6 +15,12 @@
 #include "python/src/small_vector.h"
 #include "python/src/utils.h"
 
+// These flags became part of the Stable ABI together with Py_buffer in 3.11.
+#if defined(Py_LIMITED_API) && Py_LIMITED_API + 0 < 0x030B0000
+#define PyBUF_READ 0x100
+#define PyBUF_WRITE 0x200
+#endif
+
 namespace mx = mlx::core;
 namespace nb = nanobind;
 using namespace nb::literals;
diff --git a/python/src/mlx_func.cpp b/python/src/mlx_func.cpp
index 9955e134f9..3bec0ccf14 100644
--- a/python/src/mlx_func.cpp
+++ b/python/src/mlx_func.cpp
@@ -49,7 +49,11 @@ PyObject* gc_func_vectorcall(
     PyObject* const* args,
     size_t nargs,
     PyObject* kwnames) {
-  return PyObject_Vectorcall(((gc_func*)self)->func, args, nargs, kwnames);
+  return nb::detail::vectorcall(((gc_func*)self)->func, args, nargs, kwnames);
+}
+
+PyObject* gc_func_call(PyObject* self, PyObject* args, PyObject* kwargs) {
+  return PyObject_Call(((gc_func*)self)->func, args, kwargs);
 }
 
 void gc_func_dealloc(PyObject* self) {
@@ -83,16 +87,22 @@ PyType_Slot gc_func_slots[] = {
     {Py_tp_getset, (void*)gc_func_getset},
     {Py_tp_getattro, (void*)gc_func_getattro},
     {Py_tp_members, (void*)gc_func_members},
-    {Py_tp_call, (void*)PyVectorcall_Call},
+    {Py_tp_call, (void*)gc_func_call},
     {Py_tp_dealloc, (void*)gc_func_dealloc},
     {0, 0}};
 
+#if defined(Py_TPFLAGS_HAVE_VECTORCALL)
+constexpr auto gc_func_vectorcall_flag = Py_TPFLAGS_HAVE_VECTORCALL;
+#else
+constexpr unsigned long gc_func_vectorcall_flag = 1UL << 11;
+#endif
+
 static PyType_Spec gc_func_spec = {
     /* .name = */ "mlx.gc_func",
     /* .basicsize = */ (int)sizeof(gc_func),
     /* .itemsize = */ 0,
     /* .flags = */ Py_TPFLAGS_DEFAULT | Py_TPFLAGS_HAVE_GC |
-        Py_TPFLAGS_HAVE_VECTORCALL,
+        gc_func_vectorcall_flag,
     /* .slots = */ gc_func_slots};
 
 static PyTypeObject* gc_func_tp = nullptr;
diff --git a/python/src/small_vector.h b/python/src/small_vector.h
index 1a07abe1b2..3744ea10f0 100644
--- a/python/src/small_vector.h
+++ b/python/src/small_vector.h
@@ -37,7 +37,7 @@ struct type_caster<::mlx::core::SmallVector<Type, Size, Alloc>> {
 
     // Will initialize 'size' and 'temp'. All return values and
     // return parameters are zero/NULL in the case of a failure.
-    PyObject** o = seq_get(src.ptr(), &size, &temp);
+    PyObject** o = NB_CALL(seq_get)(src.ptr(), &size, &temp);
 
     value.clear();
     value.reserve(size);
diff --git a/setup.py b/setup.py
index 52cc4f7492..853c2d985d 100644
--- a/setup.py
+++ b/setup.py
@@ -6,6 +6,7 @@
 import re
 import subprocess
 import sys
+import sysconfig
 from functools import partial
 from pathlib import Path
 
@@ -71,6 +72,8 @@ def get_version():
 build_backend = int(os.environ.get("MLX_BUILD_BACKEND_PACKAGE", 0))
 build_macos = platform.system() == "Darwin"
 build_cuda = "MLX_BUILD_CUDA=ON" in os.environ.get("CMAKE_ARGS", "")
+# Free-threaded wheels use the interpreter-specific linked ABI.
+free_threaded = bool(sysconfig.get_config_var("Py_GIL_DISABLED"))
 
 
 # A CMakeExtension needs a sourcedir instead of a file list.
@@ -78,7 +81,7 @@ def get_version():
 # If you need multiple extensions, see scikit-build.
 class CMakeExtension(Extension):
     def __init__(self, name: str, sourcedir: str = "") -> None:
-        super().__init__(name, sources=[])
+        super().__init__(name, sources=[], py_limited_api=not free_threaded)
         self.sourcedir = os.fspath(Path(sourcedir).resolve())
 
 
@@ -208,6 +211,12 @@ def get_tag(self) -> tuple[str, str, str]:
         if build_backend:
             impl = self.python_tag
             abi = "none"
+        elif not free_threaded:
+            # nanobind split mode targets the Python 3.10 Stable ABI. The
+            # official backend has no CPython 3.10 wheel for Windows ARM64, so
+            # that platform starts at CPython 3.11 instead.
+            impl = "cp311" if plat_name == "win_arm64" else "cp310"
+            abi = "abi3"
         return (impl, abi, plat_name)
 
     def write_wheelfile(self, *args, **kwargs) -> None:
@@ -299,6 +308,8 @@ def is_backend_file(file):
     install_requires = []
 
     if not build_backend:
+        if not free_threaded:
+            install_requires.append("nanobind-backend>=1.0.0.dev2")
         if build_frontend:
             install_requires.append(
                 f'mlx-metal=={version}; platform_system == "Darwin"'

From b7941714f6ffae5182ab16ffd2979714fdc89219 Mon Sep 17 00:00:00 2001
From: XXXXRT666 <157766680+XXXXRT666@users.noreply.github.com>
Date: Thu, 20 Aug 2026 02:29:11 +0800
Subject: [PATCH 3/6] Update nanobind to 3.0.0-dev3

---
 CMakeLists.txt                       |  2 +-
 _build_backend/backend.py            |  2 +-
 docs/src/dev/extensions.rst          |  2 +-
 examples/extensions/pyproject.toml   |  4 ++--
 examples/extensions/requirements.txt |  4 ++--
 examples/extensions/setup.py         |  2 +-
 python/src/convert.h                 | 18 ------------------
 setup.py                             |  2 +-
 8 files changed, 9 insertions(+), 27 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 3da61b3c47..c68f13f290 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -411,7 +411,7 @@ if(MLX_BUILD_PYTHON_BINDINGS)
   FetchContent_Declare(
     nanobind
     GIT_REPOSITORY https://github.com/wjakob/nanobind.git
-    GIT_TAG v3.0.0-dev2
+    GIT_TAG v3.0.0-dev3
     GIT_SHALLOW TRUE
     EXCLUDE_FROM_ALL)
   FetchContent_MakeAvailable(nanobind)
diff --git a/_build_backend/backend.py b/_build_backend/backend.py
index d4113225d1..60a4e92b50 100644
--- a/_build_backend/backend.py
+++ b/_build_backend/backend.py
@@ -9,7 +9,7 @@ def _with_nanobind_backend(requires):
     build_backend = int(os.environ.get("MLX_BUILD_BACKEND_PACKAGE", 0))
     free_threaded = bool(sysconfig.get_config_var("Py_GIL_DISABLED"))
     if not build_backend and not free_threaded:
-        requires.append("nanobind-backend>=1.0.0.dev2")
+        requires.append("nanobind-backend>=1.0.0.dev3")
     return requires
 
 
diff --git a/docs/src/dev/extensions.rst b/docs/src/dev/extensions.rst
index 182408dcb1..2613124ce6 100644
--- a/docs/src/dev/extensions.rst
+++ b/docs/src/dev/extensions.rst
@@ -710,7 +710,7 @@ build utilities defined in :mod:`mlx.extension`:
             cmdclass={"build_ext": extension.CMakeBuild},
             packages=["mlx_sample_extensions"],
             package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-            install_requires=["nanobind-backend>=1.0.0.dev2"],
+            install_requires=["nanobind-backend>=1.0.0.dev3"],
             options={"bdist_wheel": {"py_limited_api": "cp310"}},
             extras_require={"dev":[]},
             zip_safe=False,
diff --git a/examples/extensions/pyproject.toml b/examples/extensions/pyproject.toml
index aa3c5ef2d5..b36881c2eb 100644
--- a/examples/extensions/pyproject.toml
+++ b/examples/extensions/pyproject.toml
@@ -3,7 +3,7 @@ requires = [
   "setuptools>=42",
   "cmake>=3.27",
   "mlx>=0.18.0",
-  "nanobind==3.0.0.dev2",
-  "nanobind-backend>=1.0.0.dev2",
+  "nanobind==3.0.0.dev3",
+  "nanobind-backend>=1.0.0.dev3",
 ]
 build-backend = "setuptools.build_meta"
diff --git a/examples/extensions/requirements.txt b/examples/extensions/requirements.txt
index 2a6b060a30..e6d977cb7d 100644
--- a/examples/extensions/requirements.txt
+++ b/examples/extensions/requirements.txt
@@ -1,5 +1,5 @@
 setuptools>=42
 cmake>=3.27
 mlx>=0.31.2
-nanobind==3.0.0.dev2
-nanobind-backend>=1.0.0.dev2
+nanobind==3.0.0.dev3
+nanobind-backend>=1.0.0.dev3
diff --git a/examples/extensions/setup.py b/examples/extensions/setup.py
index 120d7050e8..4d7914acee 100644
--- a/examples/extensions/setup.py
+++ b/examples/extensions/setup.py
@@ -13,7 +13,7 @@
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-        install_requires=["nanobind-backend>=1.0.0.dev2"],
+        install_requires=["nanobind-backend>=1.0.0.dev3"],
         options={"bdist_wheel": {"py_limited_api": "cp310"}},
         zip_safe=False,
         python_requires=">=3.10",
diff --git a/python/src/convert.h b/python/src/convert.h
index 133f443ed7..7c767a5944 100644
--- a/python/src/convert.h
+++ b/python/src/convert.h
@@ -14,24 +14,6 @@ namespace nb = nanobind;
 
 namespace nanobind {
 
-template <>
-struct ndarray_traits<mx::float16_t> {
-  static constexpr bool is_complex = false;
-  static constexpr bool is_float = true;
-  static constexpr bool is_bool = false;
-  static constexpr bool is_int = false;
-  static constexpr bool is_signed = true;
-};
-
-template <>
-struct ndarray_traits<mx::bfloat16_t> {
-  static constexpr bool is_complex = false;
-  static constexpr bool is_float = true;
-  static constexpr bool is_bool = false;
-  static constexpr bool is_int = false;
-  static constexpr bool is_signed = true;
-};
-
 namespace detail {
 
 template <>
diff --git a/setup.py b/setup.py
index 853c2d985d..5feaef3962 100644
--- a/setup.py
+++ b/setup.py
@@ -309,7 +309,7 @@ def is_backend_file(file):
 
     if not build_backend:
         if not free_threaded:
-            install_requires.append("nanobind-backend>=1.0.0.dev2")
+            install_requires.append("nanobind-backend>=1.0.0.dev3")
         if build_frontend:
             install_requires.append(
                 f'mlx-metal=={version}; platform_system == "Darwin"'

From 2cf5e9c188dbb9fadd223aff3e46882b22295234 Mon Sep 17 00:00:00 2001
From: XXXXRT666 <157766680+XXXXRT666@users.noreply.github.com>
Date: Fri, 21 Aug 2026 01:29:17 +0800
Subject: [PATCH 4/6] Update nanobind to 3.0.0-dev4

---
 CMakeLists.txt                       | 2 +-
 _build_backend/backend.py            | 2 +-
 docs/src/dev/extensions.rst          | 2 +-
 examples/extensions/pyproject.toml   | 4 ++--
 examples/extensions/requirements.txt | 4 ++--
 examples/extensions/setup.py         | 2 +-
 setup.py                             | 2 +-
 7 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index c68f13f290..0558dbd679 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -411,7 +411,7 @@ if(MLX_BUILD_PYTHON_BINDINGS)
   FetchContent_Declare(
     nanobind
     GIT_REPOSITORY https://github.com/wjakob/nanobind.git
-    GIT_TAG v3.0.0-dev3
+    GIT_TAG v3.0.0-dev4
     GIT_SHALLOW TRUE
     EXCLUDE_FROM_ALL)
   FetchContent_MakeAvailable(nanobind)
diff --git a/_build_backend/backend.py b/_build_backend/backend.py
index 60a4e92b50..9c9bbedfe3 100644
--- a/_build_backend/backend.py
+++ b/_build_backend/backend.py
@@ -9,7 +9,7 @@ def _with_nanobind_backend(requires):
     build_backend = int(os.environ.get("MLX_BUILD_BACKEND_PACKAGE", 0))
     free_threaded = bool(sysconfig.get_config_var("Py_GIL_DISABLED"))
     if not build_backend and not free_threaded:
-        requires.append("nanobind-backend>=1.0.0.dev3")
+        requires.append("nanobind-backend>=1.0.0.dev4")
     return requires
 
 
diff --git a/docs/src/dev/extensions.rst b/docs/src/dev/extensions.rst
index 2613124ce6..1b3c9562f4 100644
--- a/docs/src/dev/extensions.rst
+++ b/docs/src/dev/extensions.rst
@@ -710,7 +710,7 @@ build utilities defined in :mod:`mlx.extension`:
             cmdclass={"build_ext": extension.CMakeBuild},
             packages=["mlx_sample_extensions"],
             package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-            install_requires=["nanobind-backend>=1.0.0.dev3"],
+            install_requires=["nanobind-backend>=1.0.0.dev4"],
             options={"bdist_wheel": {"py_limited_api": "cp310"}},
             extras_require={"dev":[]},
             zip_safe=False,
diff --git a/examples/extensions/pyproject.toml b/examples/extensions/pyproject.toml
index b36881c2eb..ef162c1deb 100644
--- a/examples/extensions/pyproject.toml
+++ b/examples/extensions/pyproject.toml
@@ -3,7 +3,7 @@ requires = [
   "setuptools>=42",
   "cmake>=3.27",
   "mlx>=0.18.0",
-  "nanobind==3.0.0.dev3",
-  "nanobind-backend>=1.0.0.dev3",
+  "nanobind==3.0.0.dev4",
+  "nanobind-backend>=1.0.0.dev4",
 ]
 build-backend = "setuptools.build_meta"
diff --git a/examples/extensions/requirements.txt b/examples/extensions/requirements.txt
index e6d977cb7d..113a138f30 100644
--- a/examples/extensions/requirements.txt
+++ b/examples/extensions/requirements.txt
@@ -1,5 +1,5 @@
 setuptools>=42
 cmake>=3.27
 mlx>=0.31.2
-nanobind==3.0.0.dev3
-nanobind-backend>=1.0.0.dev3
+nanobind==3.0.0.dev4
+nanobind-backend>=1.0.0.dev4
diff --git a/examples/extensions/setup.py b/examples/extensions/setup.py
index 4d7914acee..0321e0b3f0 100644
--- a/examples/extensions/setup.py
+++ b/examples/extensions/setup.py
@@ -13,7 +13,7 @@
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-        install_requires=["nanobind-backend>=1.0.0.dev3"],
+        install_requires=["nanobind-backend>=1.0.0.dev4"],
         options={"bdist_wheel": {"py_limited_api": "cp310"}},
         zip_safe=False,
         python_requires=">=3.10",
diff --git a/setup.py b/setup.py
index 5feaef3962..de96d215d2 100644
--- a/setup.py
+++ b/setup.py
@@ -309,7 +309,7 @@ def is_backend_file(file):
 
     if not build_backend:
         if not free_threaded:
-            install_requires.append("nanobind-backend>=1.0.0.dev3")
+            install_requires.append("nanobind-backend>=1.0.0.dev4")
         if build_frontend:
             install_requires.append(
                 f'mlx-metal=={version}; platform_system == "Darwin"'

From 9f674d315c75cc1510f908d839c782653a96ef8e Mon Sep 17 00:00:00 2001
From: XXXXRT666 <157766680+XXXXRT666@users.noreply.github.com>
Date: Sat, 22 Aug 2026 03:30:55 +0800
Subject: [PATCH 5/6] Update nanobind to 3.0.0-dev5

---
 CMakeLists.txt                       | 2 +-
 _build_backend/backend.py            | 2 +-
 docs/src/dev/extensions.rst          | 2 +-
 examples/extensions/pyproject.toml   | 4 ++--
 examples/extensions/requirements.txt | 4 ++--
 examples/extensions/setup.py         | 2 +-
 setup.py                             | 2 +-
 7 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 0558dbd679..7391a72093 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -411,7 +411,7 @@ if(MLX_BUILD_PYTHON_BINDINGS)
   FetchContent_Declare(
     nanobind
     GIT_REPOSITORY https://github.com/wjakob/nanobind.git
-    GIT_TAG v3.0.0-dev4
+    GIT_TAG v3.0.0-dev5
     GIT_SHALLOW TRUE
     EXCLUDE_FROM_ALL)
   FetchContent_MakeAvailable(nanobind)
diff --git a/_build_backend/backend.py b/_build_backend/backend.py
index 9c9bbedfe3..c9cc08e9a3 100644
--- a/_build_backend/backend.py
+++ b/_build_backend/backend.py
@@ -9,7 +9,7 @@ def _with_nanobind_backend(requires):
     build_backend = int(os.environ.get("MLX_BUILD_BACKEND_PACKAGE", 0))
     free_threaded = bool(sysconfig.get_config_var("Py_GIL_DISABLED"))
     if not build_backend and not free_threaded:
-        requires.append("nanobind-backend>=1.0.0.dev4")
+        requires.append("nanobind-backend>=1.0.0.dev5")
     return requires
 
 
diff --git a/docs/src/dev/extensions.rst b/docs/src/dev/extensions.rst
index 1b3c9562f4..c56fc32d35 100644
--- a/docs/src/dev/extensions.rst
+++ b/docs/src/dev/extensions.rst
@@ -710,7 +710,7 @@ build utilities defined in :mod:`mlx.extension`:
             cmdclass={"build_ext": extension.CMakeBuild},
             packages=["mlx_sample_extensions"],
             package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-            install_requires=["nanobind-backend>=1.0.0.dev4"],
+            install_requires=["nanobind-backend>=1.0.0.dev5"],
             options={"bdist_wheel": {"py_limited_api": "cp310"}},
             extras_require={"dev":[]},
             zip_safe=False,
diff --git a/examples/extensions/pyproject.toml b/examples/extensions/pyproject.toml
index ef162c1deb..894eff193f 100644
--- a/examples/extensions/pyproject.toml
+++ b/examples/extensions/pyproject.toml
@@ -3,7 +3,7 @@ requires = [
   "setuptools>=42",
   "cmake>=3.27",
   "mlx>=0.18.0",
-  "nanobind==3.0.0.dev4",
-  "nanobind-backend>=1.0.0.dev4",
+  "nanobind==3.0.0.dev5",
+  "nanobind-backend>=1.0.0.dev5",
 ]
 build-backend = "setuptools.build_meta"
diff --git a/examples/extensions/requirements.txt b/examples/extensions/requirements.txt
index 113a138f30..40f84f5c98 100644
--- a/examples/extensions/requirements.txt
+++ b/examples/extensions/requirements.txt
@@ -1,5 +1,5 @@
 setuptools>=42
 cmake>=3.27
 mlx>=0.31.2
-nanobind==3.0.0.dev4
-nanobind-backend>=1.0.0.dev4
+nanobind==3.0.0.dev5
+nanobind-backend>=1.0.0.dev5
diff --git a/examples/extensions/setup.py b/examples/extensions/setup.py
index 0321e0b3f0..cb0716c177 100644
--- a/examples/extensions/setup.py
+++ b/examples/extensions/setup.py
@@ -13,7 +13,7 @@
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-        install_requires=["nanobind-backend>=1.0.0.dev4"],
+        install_requires=["nanobind-backend>=1.0.0.dev5"],
         options={"bdist_wheel": {"py_limited_api": "cp310"}},
         zip_safe=False,
         python_requires=">=3.10",
diff --git a/setup.py b/setup.py
index de96d215d2..77853b046a 100644
--- a/setup.py
+++ b/setup.py
@@ -309,7 +309,7 @@ def is_backend_file(file):
 
     if not build_backend:
         if not free_threaded:
-            install_requires.append("nanobind-backend>=1.0.0.dev4")
+            install_requires.append("nanobind-backend>=1.0.0.dev5")
         if build_frontend:
             install_requires.append(
                 f'mlx-metal=={version}; platform_system == "Darwin"'

From 5453ad060e9a2e359b80eac2921e2d0c8e3184c2 Mon Sep 17 00:00:00 2001
From: XXXXRT666 <157766680+XXXXRT666@users.noreply.github.com>
Date: Mon, 24 Aug 2026 02:53:56 +0800
Subject: [PATCH 6/6] Update nanobind to 3.0.0

---
 CMakeLists.txt                       | 2 +-
 _build_backend/backend.py            | 2 +-
 docs/src/dev/extensions.rst          | 2 +-
 examples/extensions/pyproject.toml   | 4 ++--
 examples/extensions/requirements.txt | 4 ++--
 examples/extensions/setup.py         | 2 +-
 setup.py                             | 2 +-
 7 files changed, 9 insertions(+), 9 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7391a72093..07d6e00d8d 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -411,7 +411,7 @@ if(MLX_BUILD_PYTHON_BINDINGS)
   FetchContent_Declare(
     nanobind
     GIT_REPOSITORY https://github.com/wjakob/nanobind.git
-    GIT_TAG v3.0.0-dev5
+    GIT_TAG v3.0.0
     GIT_SHALLOW TRUE
     EXCLUDE_FROM_ALL)
   FetchContent_MakeAvailable(nanobind)
diff --git a/_build_backend/backend.py b/_build_backend/backend.py
index c9cc08e9a3..751f612780 100644
--- a/_build_backend/backend.py
+++ b/_build_backend/backend.py
@@ -9,7 +9,7 @@ def _with_nanobind_backend(requires):
     build_backend = int(os.environ.get("MLX_BUILD_BACKEND_PACKAGE", 0))
     free_threaded = bool(sysconfig.get_config_var("Py_GIL_DISABLED"))
     if not build_backend and not free_threaded:
-        requires.append("nanobind-backend>=1.0.0.dev5")
+        requires.append("nanobind-backend>=1.0.0")
     return requires
 
 
diff --git a/docs/src/dev/extensions.rst b/docs/src/dev/extensions.rst
index c56fc32d35..4722e886a7 100644
--- a/docs/src/dev/extensions.rst
+++ b/docs/src/dev/extensions.rst
@@ -710,7 +710,7 @@ build utilities defined in :mod:`mlx.extension`:
             cmdclass={"build_ext": extension.CMakeBuild},
             packages=["mlx_sample_extensions"],
             package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-            install_requires=["nanobind-backend>=1.0.0.dev5"],
+            install_requires=["nanobind-backend>=1.0.0"],
             options={"bdist_wheel": {"py_limited_api": "cp310"}},
             extras_require={"dev":[]},
             zip_safe=False,
diff --git a/examples/extensions/pyproject.toml b/examples/extensions/pyproject.toml
index 894eff193f..04185bc5e7 100644
--- a/examples/extensions/pyproject.toml
+++ b/examples/extensions/pyproject.toml
@@ -3,7 +3,7 @@ requires = [
   "setuptools>=42",
   "cmake>=3.27",
   "mlx>=0.18.0",
-  "nanobind==3.0.0.dev5",
-  "nanobind-backend>=1.0.0.dev5",
+  "nanobind==3.0.0",
+  "nanobind-backend>=1.0.0",
 ]
 build-backend = "setuptools.build_meta"
diff --git a/examples/extensions/requirements.txt b/examples/extensions/requirements.txt
index 40f84f5c98..635b14ac28 100644
--- a/examples/extensions/requirements.txt
+++ b/examples/extensions/requirements.txt
@@ -1,5 +1,5 @@
 setuptools>=42
 cmake>=3.27
 mlx>=0.31.2
-nanobind==3.0.0.dev5
-nanobind-backend>=1.0.0.dev5
+nanobind==3.0.0
+nanobind-backend>=1.0.0
diff --git a/examples/extensions/setup.py b/examples/extensions/setup.py
index cb0716c177..4bffae6d5c 100644
--- a/examples/extensions/setup.py
+++ b/examples/extensions/setup.py
@@ -13,7 +13,7 @@
         cmdclass={"build_ext": extension.CMakeBuild},
         packages=["mlx_sample_extensions"],
         package_data={"mlx_sample_extensions": ["*.so", "*.dylib", "*.metallib"]},
-        install_requires=["nanobind-backend>=1.0.0.dev5"],
+        install_requires=["nanobind-backend>=1.0.0"],
         options={"bdist_wheel": {"py_limited_api": "cp310"}},
         zip_safe=False,
         python_requires=">=3.10",
diff --git a/setup.py b/setup.py
index 77853b046a..3aab23d594 100644
--- a/setup.py
+++ b/setup.py
@@ -309,7 +309,7 @@ def is_backend_file(file):
 
     if not build_backend:
         if not free_threaded:
-            install_requires.append("nanobind-backend>=1.0.0.dev5")
+            install_requires.append("nanobind-backend>=1.0.0")
         if build_frontend:
             install_requires.append(
                 f'mlx-metal=={version}; platform_system == "Darwin"'