Metal
Apple's GPU API — the layer every Mac inference stack ultimately sits on
Apple
In plain English
The language programs use to talk to the graphics chip inside a Mac. Everything that runs AI fast on a Mac is speaking it underneath.
Technically
Apple's low-level graphics and compute API, with Metal Performance Shaders (MPS) and MPSGraph providing optimized primitives for matrix multiplication and attention. llama.cpp ships Metal kernels, MLX is built directly on it, and PyTorch reaches it through the MPS backend. Because Apple silicon shares one memory pool between CPU and GPU, Metal kernels read model weights without the host-to-device copy a discrete GPU requires — the main structural advantage of Apple hardware for inference, alongside the large memory configurations.