Rust bindings for Apple's MLX framework.

Platform: Apple Silicon (macOS) only.

cargo add mlxcore

use mlxcore::{Array, Stream};

fn main() -> mlxcore::Result<()> {
    let s = Stream::default();

    // The `f32` suffix matters — see the note further down.
    let x = Array::from_slice(&[1.0f32, 4.0, 9.0], &[3]);

    println!("{:?}", x.sqrt(&s)?.to_vec::<f32>()); // [1.0, 2.0, 3.0]
    println!("{:?}", (&x * 2.0f32).to_vec::<f32>()); // [2.0, 8.0, 18.0]

    Ok(())
}

The bindings are layered, the same way as MLX Swift:

mlxcore      (safe, idiomatic Rust API)   <- crates/mlxcore
mlxcore-sys  (raw unsafe FFI bindings)    <- crates/mlxcore-sys, generated by bindgen
mlx-c        (Apple's C API for MLX)      <- crates/mlxcore-sys/third_party/mlx-c (git submodule)
mlx          (Apple's C++ framework)      <- fetched by mlx-c's CMake build

MLX itself is C++, which Rust cannot bind to directly. We bind against mlx-c, Apple's official C API, whose CMake build pulls in MLX via FetchContent.

The mlx-c submodule sits inside mlxcore-sys rather than at the repository root because cargo package only ships files under the crate directory — from the root it would be missing from the published crate.

Within mlxcore, the modules follow the mlx-c headers they wrap:

git submodule update --init --recursive
make build
make test

Requirements: a recent Rust toolchain, make, cmake, and Xcode command-line tools. The first build compiles MLX from source and takes several minutes.

Runnable examples live in crates/mlxcore/examples:

cargo run --example hello       # arrays, shapes, streams
cargo run --example relu        # y = relu(x @ W + b) with random weights
cargo run --example attention   # a transformer attention block: QKV, RoPE, masked SDPA

Unsuffixed float literals are f64 in Rust, and Apple GPUs have no float64 support. Since the default stream is the GPU, &[1.0, 2.0] builds an array that fails on every operation. Write &[1.0f32, 2.0], and &a * 2.0f32 for scalars.

float64 still works on an explicit Stream::cpu().

• metal (default) — GPU backend via Metal.
• accelerate (default) — CPU BLAS via Apple's Accelerate framework.