deepseek-v4.1IMAGE
DeepSeek V4.1 Flash 1017 量化版双 DGX Spark 实测 67.5 tok/s
Mia-AiLab 的 EAQ v1（2.9 bpw EXL3）量化版在 2× DGX Spark、TP2、MTP 开启下单流生成中位数 67.5 tok/s（greedy），temp 1.0 为 72.5 tok/s。
Thomas Eastham
@easthambuilds11 天前收录原推发布于 2026.10.20
Thomas Eastham @easthambuilds原推发布于 2026.10.20
First speed numbers for Mia-AiLab's new DeepSeek-V4.1-Flash-1017 quant (EAQ v1, 2.9 bpw EXL3). 2× DGX Spark (TP2), MTP on, 1 stream, GPU clocks capped at 2200 MHz. Chat, 22-token prompt, 256 out (median of 5): 67.5 tok/s greedy, 72.5 at temp 1.0 (runs 65–77).
x.com/easthambuilds/status/2112239353535840338采集于 2026.10.20 12:05