2026-11-12 · America/Los_Angeles · 社区动态 · #7

KQuant - new quantization technique, 3.7× better quants for Qwen Flash Next 180B, Qwen 27B, 35B, 9B (measured in KL-divergence reduction) (MLX)

hey folks, im Alish, an ml engineer from Nace (folks behind Drex, an open source decision family model that hit #1 in Artificial analysis in decision index 2.1). Recently my cto gave me task to optimize quants for qwen4, so i dialed my uni proffesor from UoT and span up my Fable 5.2 and we cooooooked. starting with mlx (gguf tommorow). It's a pareto frontier on quantization for accuracy/size for qwen family models. Why we made these we have open sourced the fastest way to run 35b/27b on a mac(Kapitsa) and we got tired of dumb quants. So we made our own KQuant . Open source Qwen packs for…

热度 45.9 / 100；排名与评分保留该期记录。