diff --git a/.github/actions/build-macos/action.yml b/.github/actions/build-macos/action.yml
deleted file mode 100644
index 7bcd4e0d09..0000000000
--- a/.github/actions/build-macos/action.yml
+++ /dev/null
@@ -1,39 +0,0 @@
-name: 'Build macOS'
-description: 'Build C++ and Python binaries for testing on macOS'
-
-inputs:
-  cmake-args:
-    description: 'The args for generating CMake project'
-    required: true
-  macos-target:
-    description: 'The target macOS version to build for'
-    required: false
-    default: '26.2'
-
-runs:
-  using: 'composite'
-  steps:
-    - name: Install dependencies
-      shell: bash
-      run: uv pip install build setuptools
-
-    - name: Build wheel
-      shell: bash
-      env:
-        DEBUG: 1
-        CMAKE_ARGS: ${{ inputs.cmake-args }}
-        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
-      run: python -m build -w
-
-    - name: Build CPP only
-      shell: bash
-      env:
-        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
-      run: |
-        if ${{ contains(inputs.cmake-args, 'CMAKE_BUILD_TYPE') }} ; then
-          cmake . -B build ${{ inputs.cmake-args }}
-        else
-          cmake . -B build ${{ inputs.cmake-args }} \
-            -DCMAKE_BUILD_TYPE=Debug
-        fi
-        cmake --build build -j $(sysctl -n hw.physicalcpu)
diff --git a/.github/actions/build/action.yml b/.github/actions/build/action.yml
index 037e0ea6bc..da77e4817c 100644
--- a/.github/actions/build/action.yml
+++ b/.github/actions/build/action.yml
@@ -1,5 +1,5 @@
 name: 'Build'
-description: 'Build C++ and Python binaries for testing on Linux and Windows'
+description: 'Build C++ and Python binaries for testing'
 
 inputs:
   cmake-args:
@@ -8,6 +8,10 @@ inputs:
   debug:
     description: 'Do debug build'
     required: true
+  macos-target:
+    description: 'The target macOS version to build for'
+    required: false
+    default: '26.2'
 
 runs:
   using: 'composite'
@@ -15,16 +19,35 @@ runs:
     - name: Install Python package
       shell: bash
       env:
-        DEBUG: ${{ inputs.debug == 'true' && 1 || 0 }}
         CMAKE_ARGS: ${{ inputs.cmake-args }}
+        DEBUG: ${{ inputs.debug == 'true' && 1 || 0 }}
+        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
       run: |
-        # Install cpu-only torch to save space.
-        uv pip install torch --torch-backend=cpu
+        if ${{ runner.os != 'macOS' }}; then
+          # Install cpu-only torch to save space.
+          uv pip install torch --torch-backend=cpu
+        fi
         uv pip install --no-build-isolation -e ".[dev]" -v
 
+    - name: Get number of CPU cores
+      id: cpu-cores
+      shell: bash
+      run: |
+        if ${{ runner.os == 'Linux' }} ; then
+          CORES=$(nproc)
+        elif ${{ runner.os == 'macOS' }}; then
+          CORES=$(sysctl -n hw.ncpu)
+        elif ${{ runner.os == 'Windows' }}; then
+          CORES=$NUMBER_OF_PROCESSORS
+        else
+          CORES=2
+        fi
+        echo "result=$CORES" >> $GITHUB_OUTPUT
+
     - name: Build CPP only
       shell: bash
       env:
+        MACOSX_DEPLOYMENT_TARGET: ${{ inputs.macos-target }}
         # The cpp build is using some extra settings to reuse the compile cache
         # generated by the python install:
         # 1. Use the same ccache options with setup.py.
@@ -36,5 +59,4 @@ runs:
         cmake . -B build/cpp/mlx ${{ inputs.cmake-args }} \
           -DBUILD_SHARED_LIBS=ON \
           -DCMAKE_BUILD_TYPE=${{ inputs.debug == 'true' && 'Debug' || 'Release' }}
-        cmake --build build/cpp/mlx \
-          -j ${{ runner.os == 'Windows' && '$NUMBER_OF_PROCESSORS' || '$(nproc)' }}
+        cmake --build build/cpp/mlx -j ${{ steps.cpu-cores.outputs.result }}
diff --git a/.github/actions/setup/action.yml b/.github/actions/setup/action.yml
index 192576f301..7439e1bfb7 100644
--- a/.github/actions/setup/action.yml
+++ b/.github/actions/setup/action.yml
@@ -49,11 +49,13 @@ runs:
       if: runner.os == 'macOS'
       shell: bash
       run: |
-        brew update
-        brew trust aws/tap # suppress warning in github actions
         brew install openmpi ninja
         xcodebuild -showComponent MetalToolchain
         sysctl -a | grep machdep.cpu
+        if ${{ runner.environment == 'self-hosted' }} ; then
+          echo DEVELOPER_DIR=/Applications/Xcode.app/Contents/Developer >> $GITHUB_ENV
+          echo SDKROOT=/Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk >> $GITHUB_ENV
+        fi
 
     - name: Setup Windows environment
       if: runner.os == 'Windows'
diff --git a/.github/actions/test-array-api/action.yml b/.github/actions/test-array-api/action.yml
new file mode 100644
index 0000000000..3a049beaf7
--- /dev/null
+++ b/.github/actions/test-array-api/action.yml
@@ -0,0 +1,42 @@
+name: 'Run Array API tests'
+description: 'Run Array API tests'
+
+runs:
+  using: 'composite'
+  steps:
+    - name: Checkout array-api-tests
+      uses: actions/checkout@v7
+      with:
+        repository: data-apis/array-api-tests
+        ref: 'ef5b39f1190d54dca0d9a12646033d2d123afdd5' # v2026.09.08
+        submodules: 'true'
+        path: 'array-api-tests'
+        persist-credentials: false
+    - name: Install dependencies
+      shell: bash
+      run: uv pip install -r array-api-tests/requirements.txt
+    - name: Run the test suite
+      shell: bash
+      env:
+        ARRAY_API_TESTS_MODULE: mlx.core
+        ARRAY_API_TESTS_SKIP_DTYPES: 'float64,complex128'
+        PYTHONWARNINGS: 'ignore::UserWarning::,ignore::DeprecationWarning::,ignore::RuntimeWarning::'
+        METAL_DEBUG_ERROR_MODE: 0
+        METAL_DEVICE_WRAPPER_TYPE: 1
+      run: |
+        cd array-api-tests
+        if ${{ runner.environment == 'self-hosted' }} ; then
+          JOBS=4
+        else
+          JOBS=2
+        fi
+        uv run pytest array_api_tests \
+          --durations=5 \
+          -n $JOBS \
+          -c pytest.ini \
+          --derandomize \
+          --disable-data-dependent-shapes \
+          --disable-deadline \
+          -o strict_xfail=False \
+          --skips-file "${GITHUB_WORKSPACE}/python/tests/xp-tests/skips.txt" \
+          --xfails-file "${GITHUB_WORKSPACE}/python/tests/xp-tests/xfail.txt"
diff --git a/.github/actions/test-macos/action.yml b/.github/actions/test-macos/action.yml
index 13c7eafa06..3e4c56b931 100644
--- a/.github/actions/test-macos/action.yml
+++ b/.github/actions/test-macos/action.yml
@@ -20,69 +20,54 @@ runs:
         METAL_DEBUG_ERROR_MODE: 0
         METAL_DEVICE_WRAPPER_TYPE: 1
       run: |
-        run_tests() {
-          for WHEEL in dist/*.whl; do
-            break
-          done
-
+        for WHEEL in dist/*.whl; do
           echo "::group::Install $WHEEL"
           uv pip install $WHEEL
           echo "::endgroup::"
+          break
+        done
 
-          echo "::group::Run CPP tests"
-          ./build/tests/tests
-          echo "::endgroup::"
-
-          echo "::group::Run distributed tests"
-          # TODO: Re-enable for GPU after openmpi updating their libprrte
-          # dependency: https://github.com/openpmix/prrte/issues/2416
-          if ${{ inputs.toolkit == 'cpu' }} ; then
-            mpirun --bind-to none -host localhost:8 -np 8 -x MLX_MPI_LIBNAME=/opt/homebrew/lib/libmpi.dylib uv run python/tests/mpi_test_distributed.py
-          fi
-          mlx.launch --verbose -n 8 uv run python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
-          if $(grep "\[WARN\]" stderr.log); then
-            echo "Distributed ring test failed"
-            exit 1
-          fi
-          echo "::endgroup::"
+        echo "::group::Run CPP tests"
+        ./build/*/mlx.core/tests/tests
+        echo "::endgroup::"
 
-          echo "::group::Run Python tests"
-          if ${{ inputs.toolkit == 'cpu' }} ; then
-            uv run pytest --durations=10 -n 4 || uv run pytest --last-failed
-          else
-            uv run pytest --durations=10
-          fi
-          echo "::endgroup::"
+        echo "::group::Run distributed tests"
+        # TODO: Re-enable for GPU after openmpi updating their libprrte
+        # dependency: https://github.com/openpmix/prrte/issues/2416
+        if ${{ inputs.toolkit == 'cpu' }} ; then
+          mpirun --bind-to none -host localhost:8 -np 8 -x \
+            MLX_MPI_LIBNAME=/opt/homebrew/lib/libmpi.dylib \
+            uv run python/tests/mpi_test_distributed.py
+        fi
+        mlx.launch --verbose -n 8 uv run python/tests/ring_test_distributed.py -v 2> >(tee -a stderr.log >&2)
+        if $(grep "\[WARN\]" stderr.log); then
+          echo "Distributed ring test failed"
+          exit 1
+        fi
+        echo "::endgroup::"
 
-          if ${{ inputs.toolkit != 'cpu' }} ; then
-            echo "::group::Build example extension"
-            pushd examples/cmake_extension
-            uv pip install -r requirements.txt
-            uv run setup.py build_ext --inplace
-            uv run test.py
-            popd
-            echo "::endgroup::"
+        echo "::group::Run Python tests"
+        if ${{ inputs.toolkit == 'cpu' || runner.environment == 'self-hosted' }} ; then
+          uv run pytest --durations=10 -n 4 || uv run pytest --last-failed
+        else
+          uv run pytest --durations=10
+        fi
+        echo "::endgroup::"
 
-            echo "::group::Build Metal extension example"
-            pushd examples/metal_extension
-            uv pip install -r requirements.txt
-            uv run setup.py build_ext --inplace
-            uv run test.py
-            popd
-            echo "::endgroup::"
-          fi
-        }
+        if ${{ inputs.toolkit != 'cpu' }} ; then
+          echo "::group::Build example extension"
+          pushd examples/cmake_extension
+          uv pip install -r requirements.txt
+          uv run setup.py build_ext --inplace
+          uv run test.py
+          popd
+          echo "::endgroup::"
 
-        if [[ -d artifact ]] ; then
-          for SUBDIR in artifact/*; do
-            rm -rf build dist
-            mv $SUBDIR/* .
-            # Downloaded binaries all lose executable permission.
-            find build -type f -exec chmod 755 {} \;
-            # Path to metallib is hard-coded so move it besides executable.
-            mv build/mlx/backend/metal/kernels/mlx.metallib build/tests
-            run_tests
-          done
-        else
-          run_tests
+          echo "::group::Build Metal extension example"
+          pushd examples/metal_extension
+          uv pip install -r requirements.txt
+          uv run setup.py build_ext --inplace
+          uv run test.py
+          popd
+          echo "::endgroup::"
         fi
diff --git a/.github/workflows/build_and_test.yml b/.github/workflows/build_and_test.yml
index bf0ea01ee1..95fce91d6c 100644
--- a/.github/workflows/build_and_test.yml
+++ b/.github/workflows/build_and_test.yml
@@ -86,14 +86,10 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        macos-target: ['14.0', '26.2']
+        macos-target: ['26.2', '27.0']
         toolkit: ['cpu', 'metal', 'jit']
     runs-on: [self-hosted, macos]
     needs: check_lint
-    env:
-      DEVELOPER_DIR: /Applications/Xcode.app/Contents/Developer
-      # xcrun picks a newer Command Line Tools SDK that ld of Xcode cannot read.
-      SDKROOT: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
     steps:
       - uses: actions/checkout@v7
       - uses: ./.github/actions/setup
@@ -104,98 +100,20 @@ jobs:
           # Use the compile cache of metal build for all builds.
           ccache-save: ${{ matrix.toolkit == 'metal' }}
           ccache-toolkit: 'metal'
-      - uses: ./.github/actions/build-macos
-        with:
-          cmake-args: ${{ steps.setup.outputs.cmake-args }}
-          macos-target: ${{ matrix.macos-target }}
-      - uses: ./.github/actions/test-macos
-        if: matrix.toolkit == 'cpu'
-        with:
-          toolkit: 'cpu'
-      - uses: actions/upload-artifact@v7
-        if: matrix.toolkit != 'cpu'
-        with:
-          name: mlx-${{ matrix.toolkit }}-macos${{ matrix.macos-target }}
-          path: |
-            dist/mlx-*.whl
-            build/mlx/backend/metal/kernels/mlx.metallib
-            build/tests/tests
-          if-no-files-found: error
-
-  mac_test:
-    name: Test macOS (${{ matrix.toolkit }})
-    if: github.repository == 'ml-explore/mlx'
-    runs-on: [self-hosted, macos]
-    needs: mac_build
-    strategy:
-      fail-fast: false
-      matrix:
-        toolkit: ['metal', 'jit']
-    env:
-      DEVELOPER_DIR: /Applications/Xcode.app/Contents/Developer
-      # xcrun picks a newer Command Line Tools SDK that ld of Xcode cannot read.
-      SDKROOT: /Applications/Xcode.app/Contents/Developer/Platforms/MacOSX.platform/Developer/SDKs/MacOSX.sdk
-    steps:
-      - uses: actions/checkout@v7
-      - uses: ./.github/actions/setup
-        with:
-          toolkit: ${{ matrix.toolkit }}
-          use-ccache: false
-      - uses: actions/download-artifact@v8
-        with:
-          path: artifact
-          pattern: mlx-${{ matrix.toolkit }}-*
-      - run: ls -lhR artifact
-      - uses: ./.github/actions/test-macos
-        with:
-          toolkit: ${{ matrix.toolkit }}
-
-  array_api_test:
-    name: Array API Tests (Metal)
-    if: github.repository == 'ml-explore/mlx'
-    needs: mac_build
-    runs-on: [self-hosted, macos]
-    steps:
-      - uses: actions/checkout@v7
-      - uses: ./.github/actions/setup
-        with:
-          toolkit: 'metal'
-          use-ccache: false
-      - uses: actions/download-artifact@v8
-        with:
-          path: artifact
-          pattern: mlx-metal-*
-      - name: Checkout array-api-tests
-        uses: actions/checkout@v7
-        with:
-          repository: data-apis/array-api-tests
-          ref: 'ef5b39f1190d54dca0d9a12646033d2d123afdd5' # v2026.09.08
-          submodules: 'true'
-          path: 'array-api-tests'
-          persist-credentials: false
-      - name: Install dependencies
-        shell: bash
-        run: |
-          uv pip install pytest-xdist -r array-api-tests/requirements.txt
-      - name: Run the test suite
+      - name: Build python package
         shell: bash
         env:
-          ARRAY_API_TESTS_MODULE: mlx.core
-          ARRAY_API_TESTS_SKIP_DTYPES: 'float64,complex128'
-          PYTHONWARNINGS: 'ignore::UserWarning::,ignore::DeprecationWarning::,ignore::RuntimeWarning::'
-          METAL_DEBUG_ERROR_MODE: 0
-          METAL_DEVICE_WRAPPER_TYPE: 1
+          CMAKE_ARGS: ${{ steps.setup.outputs.cmake-args }} -DMLX_BUILD_TESTS=ON
+          DEBUG: 1
+          MACOSX_DEPLOYMENT_TARGET: ${{ matrix.macos-target }}
         run: |
-          if [[ -d artifact ]] ; then
-            for SUBDIR in artifact/*; do
-              rm -rf build dist
-              mv "$SUBDIR"/* .
-              break
-            done
-          fi
-          uv pip install dist/*.whl
-          cd "${GITHUB_WORKSPACE}/array-api-tests"
-          pytest array_api_tests -v -c pytest.ini -n 2 --max-examples=1000 --derandomize --disable-data-dependent-shapes --disable-deadline -o strict_xfail=False --xfails-file "${GITHUB_WORKSPACE}/python/tests/xp-tests/xfail.txt" --skips-file "${GITHUB_WORKSPACE}/python/tests/xp-tests/skips.txt"
+          uv pip install build setuptools
+          uv run python -m build -w
+      - uses: ./.github/actions/test-macos
+        with:
+          toolkit: ${{ matrix.toolkit }}
+      - uses: ./.github/actions/test-array-api
+        if: matrix.toolkit == 'metal'
 
   build_documentation:
     name: Build Documentation
diff --git a/.github/workflows/release.yml b/.github/workflows/release.yml
index 0a03b429ac..56beb9d982 100644
--- a/.github/workflows/release.yml
+++ b/.github/workflows/release.yml
@@ -141,7 +141,7 @@ jobs:
     strategy:
       matrix:
         python-version: *pyver
-    runs-on: 'macos-26-xlarge'
+    runs-on: [self-hosted, macos]
     env: *build-env
     steps:
       - uses: actions/checkout@v7
@@ -170,6 +170,12 @@ jobs:
           macos-target: '26.2'
           cmake-args: ${{ steps.setup.outputs.cmake-args }}
           build-backend: ${{ matrix.python-version == '3.10' }}
+      - name: Build macOS 27 package
+        uses: ./.github/actions/build-wheel
+        with:
+          macos-target: '27.0'
+          cmake-args: ${{ steps.setup.outputs.cmake-args }}
+          build-backend: ${{ matrix.python-version == '3.10' }}
       - name: Display content of the wheels
         run: for WHEEL in wheelhouse/*.whl; do unzip -l $WHEEL; echo; done
       - name: Upload frontend packages
@@ -214,7 +220,7 @@ jobs:
     runs-on: |-
       ${{ case(matrix.os == 'Windows', case(matrix.arch == 'aarch64', 'windows-11-arm',
                                             'windows-2022'),
-               matrix.os == 'macOS', 'macos-26-xlarge',
+               matrix.os == 'macOS', fromJson('["self-hosted","macos"]'),
                case(matrix.arch == 'x86_64' && startsWith(matrix.toolkit, 'cuda'), 'gpu-t4-4-core',
                     matrix.arch == 'aarch64', 'ubuntu-22.04-arm',
                     'ubuntu-22.04'))