mference

Swift + Metal MoE inference for Apple Silicon: Gemma 4 26B in ~2 GB, Qwen 3.6 35B in ~1.45 GB, and DeepSeek-V4-Flash 284B at up to 4.8 tok/s on a 24 GB Mac. SSD-streamed experts, native Mac app, CLI, and OpenAI-compatible server.