MLX runs Qwen 3.5 up to 2x faster than Ollama on Apple Silicon. Head-to-head benchmarks on M1 through M4, with setup instructions for both.