deepseek-v4.1POST
3 台 DGX Spark 跑 V4.1 Flash EXL3 3.75bpw：代码 106 tok/s
作者公开 3.75bpw EXL3 量化版本，在 TensorFold TP3 三卡 DGX Spark 上（频率锁 2000MHz）测得代码 106 tok/s、散文 73 tok/s；此前 2.9bpw 双卡为 88/61。
Thomas Eastham
@easthambuilds9 天前收录原推发布于 2026.10.23
Thomas Eastham @easthambuilds原推发布于 2026.10.23
@JamesTervit @antirez Nice matchup! If you want a 3-Spark column: my public 3.75 bpw EXL3 build on TensorFold TP3 runs 106 code / 73 prose tok/s in rigmark (clocks capped 2,000 MHz). Our 2.9 bpw did 88 / 61 on 2 Sparks. 🚀 Recipe: https://github.com/easthambuilds/DeepSeek-V4.1-Flash-1017-EXL3-3.75bpw-3x-DGX-Sparks-TensorFold
x.com/easthambuilds/status/2113464078500778208采集于 2026.10.23 12:06