deepseek-v4.1POST

GB10 上 1 到 8 并发：V4.1 Flash 67→140.9 tok/s

同机同引擎从 1 路扩到 8 路对话，DeepSeek V4.1 Flash 吞吐 67 涨到 140.9 tok/s，GLM 5.3 Flash 为 88.1→177.2，作者问家用机该按几路来配。

CryptoBandit

@CryptoBandit1009 天前收录原推发布于 2026.10.23

CryptoBandit @CryptoBandit100原推发布于 2026.10.23

Same 4 GB10 boxes, same engine, scaling from 1 to 8 chats: GLM 5.3 Flash 88.1 to 177.2 tok/s DeepSeek V4.1 Flash 67 to 140.9 tok/s Temp 0, thinking off, 512 tokens, median of 3. GLM adds more per chat too. Do you size a home rig for one chat or for many?

x.com/cryptobandit100/status/2113389938834366947采集于 2026.10.23 12:05