Nace.ai releases Kapitsa, an open-source Apple Silicon inference engine for Qwen and Nemotron

Details

Nace.ai released Kapitsa, an open-source (Apache 2.0) inference engine optimized for Apple Silicon to run Qwen and NVIDIA Nemotron models locally.
On a MacBook Pro M4 Pro (48GB), Kapitsa achieves 154 tok/s for Qwen3.6 35B-A3B and 63 tok/s for Qwen3.8 27B, surpassing mlx-serve, Ollama, and llama.cpp.
On a MacBook Pro M5 Max (128GB), the engine reaches 272 tok/s for Qwen3.6 35B-A3B and 105 tok/s for Qwen3.8 Flash-Next.
Performance gains come from custom Metal kernels per architecture, speculative decoding (DFlash2, DSpark, LiLiCorr), and hybrid KV/recurrent-state caching.
Supports Qwen 3.5 4B, 3.8 27B, 3.6 35B-A3B, 3.8 Flash-Next, and Nemotron 3.5 Lightning, Bolt, Super-VL, and Dory 4B.
Requires Apple Silicon (M1 or newer) and macOS 14+; other model architectures are not supported.

Key point