Apple Silicon and macOS VMs: 11–16× Faster LLM Inference with Llama.cpp

Researchers achieved 11–16× faster large language model (LLM) inference performance on Apple Silicon macOS virtual machines by implementing a compatibility layer that unlocks newer Metal graphics paths in the guest OS. Testing with TinyLlama and Google's Gemma models showed the optimized VMs reached 94–99% of bare-metal performance while maintaining full virtualization, addressing a significant limitation in macOS VM graphics capabilities for AI workloads.