Primary sourcearXiv cs.AI
View sourceEvent ·Single source
Gated DeltaNet在4位量化中保持性能的机制研究
Minima研究团队对Qwen3.8-27B混合模型进行了全NVFP4 W4A4量化,在4K/32K困惑度、MMLU-Pro、GSM8K等7项基准上与BF16精度相当。该模型仅17.5 GiB,预填充速度提升14-19%,32K困惑度差距随位置增加而缩小。研究揭示了四个关键机制:NVFP4的16元素块缩放局部化残差流极值异常,门控投影对误差不敏感,delta规则递归在32K token内保持噪声平坦,以及每token量化成本随上下文增长而非累积。
Current read
Minima团队把Qwen3.8-27B全量化到4位,性能不变还更小更快,揭秘了混合模型递归部分为什么好量化。
2 sources73° AI heatLast updated 2026/9/3 17:04:26
Evidence
2 sourcesRelated source 1Hugging Face: Blog
View sourceConflict check
Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.