GitHub radar
Run Kimi K3 Locally on Mac: Deltafin
Deltafin lets you run Kimi K3, a 2.8-trillion-parameter AI model by Moonshot AI, on a single Apple Silicon Mac with an OpenAI-compatible API server.
Deltafin is a research project that runs Kimi K3 — Moonshot AI's 2.8-trillion-parameter Mixture-of-Experts LLM — on a single Apple Silicon Mac or Linux machine, streaming model experts on demand from a local disk cache. Kimi's K2.x line (which K3 builds on) uses a DeepSeek V3-compatible architecture; deltafin's research notes compare the inference approach against DeepSeek V4 models. It includes an OpenAI-compatible API server, so any app that talks to ChatGPT can be pointed at it instead. On an M1 Max with 64 GB RAM, it generates roughly one token every 15 seconds. Full local mode needs about 1.7 TB of disk; a streaming mode works with 215 GB at the cost of slower inference.
Why a vibe-coder should care
Kimi K3 is one of the largest publicly available AI models; deltafin makes it possible to run it without any cloud subscription or API key. The tradeoff is real — you need a capable Mac, lots of disk, and patience — but it demonstrates that frontier-scale AI can run entirely on personal hardware today.
▌ More finds