MLX

Apple's machine learning research team built MLX as an array framework designed specifically for Apple silicon, and it has become the standard base for local LLM inference and fine-tuning on Macs. Arrays live in unified memory, so operations run on CPU or GPU without copying data between devices, and computation is lazy: graphs are built dynamically, materialized only when needed, and never trigger slow recompilation when input shapes change. The API mirrors NumPy with higher-level neural network modules, composable function transformations provide automatic differentiation, vectorization, and graph optimization, and bindings cover Python, C++, C, and Swift. Installation is pip install mlx on macOS, with CUDA and CPU-only variants for Linux. The wider ecosystem matters as much as the core: mlx-lm handles LLM generation and fine-tuning, mlx-swift brings models to iOS, and mlx-examples covers Stable Diffusion and Whisper. MIT licensed with about 27.8k GitHub stars, it is the reason a consumer MacBook can fine-tune and serve open-weight models locally.