deepseek-v4.1POST
单卡 3090 + 64GB 内存跑 V4.1 Flash 的三值化设想
作者提出把 V4.1 Flash 1017 三值化后塞进单张 RTX 3090 + 64GB 内存，正在研究如何在压缩后尽量不损失 agent 精度，尚未给出实测数据。
Casper Hansen
@casper_hansen_10 天前收录原推发布于 2026.10.17
Casper Hansen @casper_hansen_原推发布于 2026.10.17
If you could run DeepSeek v4.1 Flash 1017 on a single RTX 3090 + 64GB RAM, would you? This should be possible by ternarizing the model, but the open problem is how to keep enough agent accuracy to not feel too much degradation. I’m looking into this btw…
x.com/casper_hansen_/status/2111422186267291830采集于 2026.10.18 16:41