MLX

MLX is Apple’s open-source framework for machine learning, built specifically for Apple silicon (M-series chips). Its headline feature is unified memory: the CPU and GPU share one pool of memory, so data doesn’t have to be copied back and forth between them the way it does with CUDA .

Under the hood it builds on Apple’s Metal GPU stack. The API is deliberately close to NumPy (with PyTorch-style building blocks for neural networks), and it only does the actual computing when you ask for a result rather than eagerly along the way (lazy evaluation). It also works out the calculus that training needs automatically, and can shrink models to smaller number formats right on the device.

For running LLMs locally on a Mac, the companion library mlx-lm loads and runs models in MLX format, the native Apple-silicon alternative to the GGUF / llama.cpp world (often noticeably faster on a Mac, but Mac-only).