Llama 3 8B hits 60 tok/s on MacBook via MLX

Apple Silicon's unified memory architecture is quietly turning MacBooks into legitimate local AI workstations — and MLX is the framework making it practical.

• –MLX eliminates CPU-GPU data transfer overhead via unified memory, giving M-series MacBooks a structural edge over discrete GPU rigs for LLM inference
• –Llama 3 8B has become the de facto benchmark model for local inference — fast enough to be useful, compact enough to fit on consumer hardware
• –MLX has closed the gap with llama.cpp and is now the recommended Apple Silicon runtime, with M4/M5 memory bandwidth gains translating directly to token throughput gains
• –Apple formally backed MLX at WWDC25 with a dedicated session, signaling this is a long-term platform investment
• –For developers who want privacy-first, offline AI, the MLX + MacBook stack is increasingly the default answer