Running Kimi K3 on a M1 Mac
Summary
Deltafin documents an experiment to run Kimi K3, a 2.8T-parameter Mixture-of-Experts LLM, on a single Apple Silicon Mac. The setup uses a resident spine on disk with on-demand fetching of 82,432 experts, int8 quantization, and an OpenAI-compatible server; full install requires ~1.7 TB, streaming ~215 GB, with substantial speedups once data is cached. It emphasizes the work as a research artifact rather than a production-ready chat system and notes dependencies and licensing.