• Every nace-ai KQuant pack now writes with its drafter: kapitsa serve --model nace-ai/Qwen3.5-9B-MLX-Q4-KQuant writes about 145 tok/s on an M5 Pro, up from about 52 (#1210, #1191).
• KQuant packs read long prompts faster on M5 Macs: a 4K-token prompt in 2.4 s instead of 3.9 s on the 9B (#1199).
• On M5 Macs, the default 9B is now our Q4-KQuant pack: answers closer to the original model, writing about 5% slower. M3 and M4 Macs keep the older pack (#1196).