今天新增minimax-m3.1IMAGE
双 DGX Spark 跑 M3.1 Flash 2.15bpw：33 tok/s，质量下滑
两台 DGX Spark 以 2.15bpw 跑 428B M3.1 Flash，改写 33 tok/s、代码 29 tok/s，8k 提示读取 13.6 秒；但 IFEval 仅 60.5%，远低于 DeepSeek V4.1 Flash 的 89.0%，正试 3.06bpw 包。
ata
@atasrb6 小时前收录原推发布于 2026.10.28
ata @atasrb原推发布于 2026.10.28
MiniMax-M3.1-Flash (428B MoE) running on my two DGX Sparks at 2.15 bpw. 33 tok/s on rewrite, 29 on code. @ViC305's EXL3 pack and vLLM EXL3 plugin, with @Tech2Wild's two-Spark patches. Half the model on each Spark over QSFP, 30 GiB still free on each. An 8k prompt reads in 13.6 s. Quality is the weak part. IFEval (thinking off) is 60.5%, DeepSeek-V4.1-Flash at 2.9 bpw got 89.0% on the same set. Most of the experts here are 2-bit, so the 3.06 bpw pack is running now on the same setup to see how much of that is the quant.
这条是引用着另一条写的：「M3.1 Flash EXL3 量化：单张 DGX Spark 跑 428B，Top-1 一致率 96.89%」
x.com/atasrb/status/2115152274578153736采集于 2026.10.28 12:02