MLX (Apple)
Apple's open-source NumPy-like array framework for machine learning on Apple silicon, with unified memory, lazy compute and mlx-lm for LLMs
Overview
MLX is "a NumPy-like array framework designed for efficient and flexible machine learning on Apple silicon," built by Apple machine learning research and released as open source (MIT) in December 2023. Arrays live in unified memory, so CPU and GPU work on the same data without copies, and computation is lazy until a result is needed. It has a Python API that tracks NumPy and a C++ API, with Swift and C APIs referenced by Apple. The companion mlx-lm package runs and fine-tunes LLMs on a Mac. On M5 chips MLX uses the GPU Neural Accelerators through Metal 4: Apple measured 3.33x to 4.06x faster time to first token than M4 and 1.19x to 1.27x faster token generation (macOS 26.2 or later). Recent releases add CUDA kernels too, a sign of a Linux CUDA backend in development, though Apple has not published a stable support statement for it. Latest release: v0.32.4 (October 2026), adding Python 3.15 support and 1-bit affine quantization on Metal.
The Verdict
Who Should Use MLX?
Best For
- Developers running or fine-tuning LLMs locally on Apple silicon Macs
- Researchers who like NumPy-style code with composable transforms
- Swift and macOS app teams adding on-device ML
Not Ideal For
- Training on NVIDIA clusters or TPUs (use PyTorch or JAX)
- Teams that need a stable 1.0 API (MLX is still 0.x)
- Developers without Apple silicon hardware
What's Great
- Unified memory: no copies between CPU and GPU
- Uses M5 Neural Accelerators for large speedups on prompt processing
- mlx-lm makes local LLM inference and fine-tuning simple
- Supported in the wider ecosystem: Transformers docs list mlx, Keras names MLX as a backend
Watch Out For
- Built for Apple silicon first; CUDA support is still in development
- Pre-1.0 versioning, so APIs can change between releases
- Token generation speed is bound by memory bandwidth, so gains there are smaller
Pricing
View all features & details
Key Features
- Unified memory model across CPU and GPU
- Lazy computation
- Python API close to NumPy, plus C++ (Swift and C referenced by Apple)
- v0.32.4: VJP for scaled dot product attention, Metal 1-bit affine quantization, Neural Accelerator path for gather_qmm
Related Projects
- mlx-lm: run and fine-tune LLMs with MLX (MIT)
- M5 benchmarks: FLUX-dev 4-bit 1024px images over 3.8x faster than M4
Real-World Usage
Apple's Own Benchmarks
- MacBook Pro M5 vs M4 (24GB): 3.33x to 4.06x faster time to first token
- Token generation 1.19x to 1.27x faster (memory bandwidth 120 to 153 GB/s)
Ecosystem
- Hugging Face Transformers docs list mlx among adjacent libraries
- keras.io names MLX among Keras backends
How It Compares
| Feature | MLX | PyTorch | JAX | Transformers |
|---|---|---|---|---|
| Focus | ML on Apple silicon | General deep learning | Accelerator array compute | Pretrained model definitions |
| Mac GPU | Yes, unified memory | MPS backend | No (CPU only on macOS) | Through PyTorch MPS |
| Versioning | 0.x | 2.x | 0.x | 5.x |
| License | MIT | BSD-style | Apache-2.0 | Apache-2.0 |
Sources: GitHub · PyTorch 2.15 · JAX install docs
Is this your tool? Add a badge