nace-ai/kapitsa v0.1.36 is an update to an engine targeting Qwen and Nemotron inference on Apple Silicon. This release consolidates per-model requirements and memory information, and refines a comparative benchmark table across models. It also addresses a server crash that occurs when the number of threads used to compile the kernel differs from the number of execution threads on M1, and improves macOS 14 build compatibility. The changes are listed in the v0.1.36 release notes on GitHub.
Kapitsa is introduced as an inference engine optimized for Apple Silicon, targeting Qwen and Nemotron. The repository description also states that it provides an interface compatible with the OpenAI and Anthropic APIs. Among these changes, the per-model requirements and memory consolidation lets users check the conditions needed before use, while the comparative benchmark table organizes material for looking at results across models. However, the facts provided contain no actual figures from the table or the measurement conditions.
What can therefore be confirmed in this release is the consolidation of documentation and tables, the server crash stemming from M1 thread settings, and the improvement to macOS 14 build compatibility. The fact that the table was tidied up is not enough to judge the superiority of any particular model or real inference speed. Performance figures and the measurement environment were not presented in the given release summary, and the stated target is also Qwen and Nemotron inference on Apple Silicon. So it is more accurate to read this as a record of changes for the stated environment rather than as a performance evaluation that extends to other hardware or models.