Kapitsa offers native Metal inference on Apple Silicon under Apache 2.0 license | Tool Update
Kapitsa is an open source inference engine licensed under the Apache 2.0 license, built by nace.ai and optimized for Apple Silicon with a native Metal implemen…
Published on MyPrivateClaw
Oct 19, 2026, 11:15 PM UTC
Coverage date
Oct 19, 2026
Last updated
Oct 19, 2026, 11:15 PM UTC
News summary
Kapitsa is an open source inference engine licensed under the Apache 2.0 license, built by nace.ai and optimized for Apple Silicon using a native Metal implementation, according to the vendor reported repository. The project states that it supports Qwen and Nemotron models with OpenAI + Anthropic API compatibility. The engine's native Metal runtime uses distilled flash attention (DFlash/DFlash2) with multi token prediction and speculative decoding, and it retains per conversation KV cache so follow up reads only new input, according to the repository. Kapitsa requires a Mac with Apple silicon (M1 or newer) and macOS 14 Sonoma or later, and installs via brew install nace ai/tap/kapitsa. On a MacBook Pro 16 inch with an M4 Pro and 48 GB, the vendor reports that Kapitsa decodes a 512 token Qwen3.6 35B A3B answer at 4 bit at 154 tokens per second, the fastest among the nine compared engines…