deepseek-v4.1POST
V4.1 Flash 1017 在 4× CMP 170HX 上 EXL3 量化：结构化 180 tok/s
本地部署实测：仅路由专家做 EXL3 2.9 量化，其余保留官方 FP8；结构化 180、代码 137、散文 97 tok/s，KV 池 1.53M，30K/121K/241K 检索 9/9。
BeyondBacktesting
@BBacktesting12 天前收录原推发布于 2026.10.18
BeyondBacktesting @BBacktesting原推发布于 2026.10.18
DeepSeek V4.1 Flash 1017 4× CMP 170HX ⚡️🔥🎙️ Structured 180 tok/s Code 137 tok/s Prose 97 tok/s Aggregate tok/s, 2 / 4 / 8: Structured 248 / 419 / 621 Code 128 / 173 / 376 Prose 134 / 188 / 283 KV pool 1.53M Routed experts only are EXL3 2.9. Attention, shared experts, head, draft, and Engram stay the official FP8 1017 weights. LibertAI's full 2.9 pack also quantizes attention: +1.8% code perplexity, +0.4% prose vs unquantized. This build quantizes less, so it should sit closer. Needle 9/9 at 30K, 121K, and 241K https://github.com/CurtisAccelerate/DeepSeekv41-Flash-1017-EXL3-CMP170HX
x.com/bbacktesting/status/2111723147527323864采集于 2026.10.18 16:33