MLX

MLX is Apple's machine-learning array framework, built by Apple ML Research for Apple silicon: a NumPy-familiar Python API (plus C/C++/Swift) with lazy computation, composable function transformations (autodiff, vectorization), and a unified memory model — arrays live in shared memory, so there are no CPU/GPU copies. It's the reason local LLM inference on Macs works so well (whisper, Stable Diffusion, and LLaMA-class models all have MLX ports), and it now ships a CUDA backend for Linux too (pip install mlx[cuda]). MIT, 28.7k stars. Researcher-oriented like JAX: friendly, but the ecosystem is smaller than PyTorch's.

```bash

pip install mlx # macOS (Apple silicon)

# pip install mlx[cuda] # Linux with NVIDIA GPU

# pip install mlx[cpu] # Linux CPU-only

```

```python

import mlx.core as mx

a = mx.array([1.0, 2.0, 3.0])

print(mx.sum(a)) # lazy — materializes on use

```

Model examples (llama, whisper, SD): github.com/ml-explore/mlx-examples. Docs: ml-explore.github.io/mlx.