Kapitsa 0.2.26 speeds up local inference on Apple Silicon

On November 2, 2026, a Reddit post in LocalLLaMA from Ye, one of the contributors to Kapitsa, announced the release of version 0.2.26 of the open-source project. Kapitsa is an inference engine built for Apple Silicon, designed to run models such as Qwen, Nemotron, Flint and Clover locally. According to Ye, the engine relies on custom-written Metal kernels that are tuned for each chip and each model, and the project is hosted at https://github.com/nace-ai/kapitsa. Compared with the team's earlier M5 Max benchmark numbers, Ye said the latest update makes Flash-Next 180B read prompts 46% faster, raises write speed for the 27B model by 8%, and adds DiffuSpace's Clover 35B to the supported model list. The announcement came through the LocalLLaMA community, where the post was shared at 06:47 on November 2, 2026, and the project's README was updated to point to 0.2.26. These speed gains are reported by the contributor rather than independently verified, and the materials available do not yet say when users might see the next update or which benchmarks will be used to measure future improvements.