今天新增deepseek-v4.1POST
双 DGX Spark 跑 V4.1 Flash：EXL3 2.80 BPW，STEM 93/100
用户分享双 DGX Spark 上 TP=2 的 DeepSeek V4.1 Flash 量化方案：EXL3 2.80 BPW、每会话 1M 上下文，自建 353 题 MMLU-Pro STEM 得 93 分。
JJ Eaton
@jayleaton8 小时前收录原推发布于 2026.10.28
JJ Eaton @jayleaton原推发布于 2026.10.28
Everyone is waking up to DeepSeek V4.1 Flash 1017. It's my favourite local model. If you have 2 DGX Sparks, my TP=2 quant and recipe have been up for over a week, and it's still very good. - EXL3, 2.80 BPW - 1M context per chat - MMLU-Pro STEM 93/100 on my 353-question set Links below.
x.com/jayleaton/status/2115271641408372950采集于 2026.10.28 12:01