davide-troiani/moeshift: Memory efficient Mixture-of-Experts for Apple Silicon

MoEShift is a memory-efficient and high-performance Mixture-of-Experts inference runtime optimized for Apple Silicon Macs. It allows users to run large models like Qwen 3.8 Flash Next locally under 10 GB of peak process memory while keeping the system responsive for other applications. The project provides setup tools via Homebrew, an OpenAI-compatible local API server, and configurable settings for expert caching and prefill behavior. Open article →