Cua Releases Metal Shim That Speeds llama.cpp in Apple Silicon macOS VMs

Cua has released an experimental Metal capability shim that it says makes llama.cpp inference in macOS virtual machines on Apple Silicon dramatically faster, without changing hardware or replacing Apple's paravirtualized GPU path. In Cua's M1 Ultra tests, Google's Gemma 4 12B QAT Q4_0 generation rose from 3.41 to 49.67 tokens per second, a 14.54x speedup, while prompt processing reached 515.76 tokens per second, or 99.59% of the bare-metal host result.