Kapitsa, an Apache 2.0 Apple Silicon inference engine, reports 128 tok/s decode for Recursive's Flint 40B-A2.5B on a MacBook Pro M4 Pro 48GB
Kapitsa, an Apache 2.0 inference engine for Apple Silicon written in Rust with custom Metal kernels and built by nace.ai, reports roughly 128 tok/s decode for…
Published on MyPrivateClaw
Oct 22, 2026, 4:42 PM UTC
Coverage date
Oct 22, 2026
Last updated
Oct 22, 2026, 4:42 PM UTC
News summary
Kapitsa is an open source inference engine for Apple Silicon written in Rust with custom Metal kernels covering M1 through M6, built by nace.ai and licensed Apache 2.0. It is OpenAI and Anthropic API compatible and runs Qwen, Nemotron and Recursive's Flint models locally to serve OpenAI and Anthropic APIs from a local model; the description is the project's own, as stated on its GitHub repository. In Kapitsa's own benchmark, Recursive's Flint 40B A2.5B in a 4 bit MLX export achieves 128 tok/s decode on a MacBook Pro M4 Pro 48GB using Kapitsa. The GitHub benchmarks record a measured 127.9 tok/s for Kapitsa versus 126.6 tok/s for Recursive's own rec 0.1.0 engine on 2026 10 22, all decoding at the model card's temperature 1.0 with top p 0.95 and top k 20. The numbers are ranked by measured value and are self reported. Kapitsa requires a Mac with Apple silicon (M1 or newer) and macOS 14 Son…