mlx-build

Unofficial precompiled binaries of the Apple MLX library, ml-explore/mlx.

Caution

This repo hosts unofficial precompiled binaries of ml-explore/mlx.

This repo is not endorsed by or has any association with the original source code. The precompiled binaries are produced on the GitHub platform and is mainly used by the Elixir library emlx.

Please evaluate and use at your own risk.

Artifacts

Precompiled binaries are published on every release; the variants and file names differ by platform.

macOS

Each release from v0.31.0 onward ships 8 macOS variants — the cross product of three axes:

axis	values	effect
build type	release / debug	debug = Debug CMake build + -D MLX_METAL_DEBUG=ON
Metal kernels	AOT / jit	jit = -D MLX_METAL_JIT=ON (kernels compiled at runtime)
deployment target	14.0 / 26.2	minimum macOS the binary runs on (see below)

All are arm64-apple-darwin, built on macos-26 (Xcode 26).

mlx-arm64-apple-darwin-<deployment-target>[-debug][-jit].tar.gz

e.g. mlx-arm64-apple-darwin-14.0.tar.gz, mlx-arm64-apple-darwin-26.2-debug-jit.tar.gz.

Naming changed at v0.31.0. The deployment-target segment is new. Releases from v0.31.0 onward use the 8-variant scheme above; earlier releases (v0.30.x and below) use the old mlx-arm64-apple-darwin[-debug][-jit].tar.gz names (4 variants, no deployment target).

The filename is keyed on the deployment target, not MLX_METAL_VERSION, because the deployment target is the decisive factor: it is the one knob we set, and everything else derives from it. MLX probes the Metal language version with -mmacosx-version-min=<deployment-target>, so the deployment target alone determines the Metal version, the minimum OS, and whether the NAX kernels are compiled in:

deployment target	MLX_METAL_VERSION	runs on	ahead-of-time NAX kernels
14.0	310 (Metal 3.1)	macOS 14+	❌ gated out — portable
26.2	400 (Metal 4.0)	macOS 26.2+	✅ baked into mlx.metallib

NAX kernels are MLX's GEMM/attention paths built on Apple's MetalPerformancePrimitives (Metal 4 tensor ops); they need the macOS 26 SDK and only run on macOS 26.2+. MLX gates them behind MLX_METAL_VERSION >= 400 AND MACOS_SDK_VERSION >= 26.2 AND CMAKE_OSX_DEPLOYMENT_TARGET >= 26.2 (kernels/CMakeLists.txt). On our macos-26 build host the SDK is always ≥ 26.2, so the deployment target alone flips NAX on or off.

Use the 14.0 builds for broad compatibility; use 26.2 only if you target macOS 26.2+ and want NAX acceleration compiled in. (The jit variants additionally JIT-compile NAX at runtime on 26.2+ hardware, regardless of deployment target.)

Linux

Two C libraries are built, each in release and debug:

mlx-<arch>-linux-<gnu|musl>[-debug].tar.gz

libc	arches
glibc (-linux-gnu)	x86_64, aarch64, riscv64, armv7l (armv7l-linux-gnueabihf), ppc64le, s390x
musl (-linux-musl)	x86_64, aarch64, riscv64, armv7l (armv7l-linux-musleabihf)

e.g. mlx-x86_64-linux-gnu.tar.gz, mlx-aarch64-linux-musl-debug.tar.gz, mlx-s390x-linux-gnu.tar.gz.

All are CPU builds with the full feature set — BLAS/LAPACK via OpenBLAS, GGUF + safetensors, distributed (ring + MPI). glibc x86_64/aarch64 target a ~2.14 floor (Ubuntu 20.04 + gcc-13); the rest build on Ubuntu 24.04 / Alpine. Each binary dynamically links OpenBLAS + libgfortran + libstdc++, so the target must provide them (apt install libopenblas0 libgfortran5 or apk add openblas libgfortran libstdc++).

Linux (CUDA)

For NVIDIA GPUs there is a separate CUDA variant (glibc, x86_64 + aarch64):

mlx-<arch>-linux-gnu-cuda13-cudnn9[-debug].tar.gz

Built against CUDA 13.0 + cuDNN 9 (MLX_BUILD_CUDA=ON — adds cuBLAS/cuFFT/cuDNN/NCCL) for compute capabilities 8.0 / 9.0 / 10.0 / 12.0 (+ PTX for forward compatibility). It needs an NVIDIA GPU with the CUDA 13 runtime + cuDNN 9 + driver at runtime, so it is a separate, non-portable variant — not a drop-in replacement for the CPU builds.

Usage

Install a release in a workflow with the bundled action — it auto-detects the target triplet:

- uses: cocoa-xu/mlx-build@main
  with:
    mlx-version: "~> 0.32"        # exact ("0.32.0") or requirement (">= 0.31.0")
    # debug: true                # debug variant
    # deployment-target: "26.2"  # macOS: NAX-enabled build (default 14.0)
    # jit: true                  # macOS: JIT Metal variant
    # cuda: "13"                 # Linux: NVIDIA CUDA 13 + cuDNN 9 variant

It exports MLX_DIR / CMAKE_PREFIX_PATH (for find_package(MLX)) and LD_LIBRARY_PATH / DYLD_LIBRARY_PATH, and sets install-dir, lib-dir, include-dir, and the resolved mlx-version / triplet as step outputs.