Apple Silicon and macOS VMs: 11–16× Faster LLM Inference with llama.cpp
Summary
The article reports on a research release that uses a per-process Metal capability shim to unlock newer Apple Metal paths in a macOS VM, boosting llama.cpp inference on Apple Silicon. Benchmarks show up to 11x prompt speed and 16x token generation improvements on an M1 Ultra; it also tests Gemma 4 12B and MLX-LM. It includes caveats about experimental, per-process scope, and provides steps to reproduce inside a Lume VM.