nace-ai/kapitsa released v0.1.38. It is an inference engine optimized for Apple Silicon, targeting Qwen and Nemotron, and is compatible with the OpenAI and Anthropic APIs. This release includes the full measurement table from an M1 Pro 16GB environment and the Qwen3.5-4B procedure. It also includes the Dory processing speed figures that were measured in a previous release. A fix for the case where a response is cut off in the middle of a character is included as well. In that case, it now returns valid UTF-8 containing replacement characters instead of broken UTF-8.
kapitsa is an inference engine that runs trained models. The repository description highlights Apple Silicon optimization and support for Qwen and Nemotron, and presents compatibility with the OpenAI and Anthropic APIs. The measurement data in this release is the full table from an M1 Pro 16GB environment plus the Qwen3.5-4B procedure. The Dory processing speed is introduced not as something newly measured in this release, but as something measured in a previous release. As a result, the release summary alone does not reveal the details of individual measurements or measurement conditions.
The UTF-8 fix is a change that returns a valid encoding even when a response is cut off in the middle of a character. However, the provided FACTS contain neither the values of the measurement table nor a description of Dory. Therefore, this material alone makes it hard to judge the speed differences between models or the meaning of the measurement results. To compare performance, you need to check the table and procedure included in the release directly.