Kapitsa

Fast inference on Apple silicon. 351 tok/s on Qwen3.6 35B-A3B on an M5 Max.

brew install nace-ai/tap/kapitsa Copied

Benchmarks

Benchmark poster: prefill and decode speed for Kapitsa and nine other engines across five Macs

Prefill / decode in tok/s, 4K prompt, 512-token answer. Full results

Which model fits my Mac?

Start in one command

Kapitsa serves OpenAI and Anthropic APIs locally. Start a coding agent on the local model:

kapitsa opencode --model 35b

Also: kapitsa claude, kapitsa codex, kapitsa pi

GitHub Docs Benchmarks Community leaderboard