MLX

MLX is Apple's machine learning framework built specifically for Apple Silicon. Unlike running llama.cpp through Metal, MLX is designed ground-up for the unified memory architecture of M-series chips, which means model weights and KV cache can be shared between CPU and GPU with no copy overhead. For local inference on a Mac, this delivers noticeably better tokens-per-second than the generic options at the same memory footprint. The ecosystem now includes mlx-lm for LLM inference with a simple Python API, mlx-vlm for vision-language models, and community-maintained quantized weights for most popular open-source LLMs. For anyone doing serious local work on a MacBook Pro or Mac Studio, MLX is the default inference layer in 2026.