The GitHub release lists a server-crash fix for systemone when a question exceeds 2,048 tokens, plus rejection of overly deep JSON before it reaches the parser.
The release reports faster prefill for a 35B-A3B mixture-of-experts model on an M5: 7.1% at 4K tokens and 6.5% at 16K tokens. Kapitsa says the resulting output is bitwise identical. It also adds kapitsa-systemone and kapitsa-drex-pointer to release packages, and changes releases to fail if the Homebrew tap cannot be updated.
Other changes include wrapping every decode with KFlash, with prompt-lookup rows opt-in through NI_KFLASH_BUDGET=1. The notes also describe a Flint decode issue producing garbage output on an M5 Ultra, not verified on a GPU, and a terminal-query leak fix. The source does not provide independent benchmarks, general performance figures across hardware, or evidence that the listed fixes have been validated outside the project’s stated tests.