nace-ai / kapitsa

What it does

• 351 tok/s on Qwen3.6 35B-A3B (M5 Max), the fastest decode in 33 of 38 model and Mac pairs. KFlash, tree speculative decoding: one verify pass checks a tree of the drafter's guesses, with a DFlash2, DFlash, MTP, DSpark or LiLiCorr drafter be
• Downloads closer to the original model, at the same size. The Qwen3.5 4B and 9B packs are KQuant builds: the same MLX 4-bit files, a third less error on the 4B and half on the 9B (KQuant).
• A 180B model on a 48 GB Mac, at 61 tok/s (M5 Pro). Its routed experts stream from the SSD, as many as the budget holds.
• A 16K follow-up after a restart starts in 0.16 s instead of 17 s. Conversations stay on the SSD (--disk-cache).
• Prefill on the GPU and the Neural Engine together. Prefill hands part of its work to the Neural Engine, which the GPU would otherwise wait on.