New /engine page shows what Kapitsa is doing (memory, slots, drafts, requests) and lets you Stop it or Switch model from the page.
Prompts are read up to 1.9x faster on M5 Macs for 3-bit, 8-bit and GGUF 4-bit models (4K tokens, Qwen3.5 4B).
KQuant 4-bit stays closer to the original model at the same size: it is the 4-bit pack for Qwen3.6 35B-A3B and Qwen3.5 4B and 9B, and an opt-in for Qwen3.8 27B (kapitsa serve --model 27b:kquant). Qwen3.8 27B also writes ~8% faster on the M5 Pro with our own drafter.