deepseek-v4.1IMAGE
RTX 5090 上 2.25bpw 量化跑 47 tok/s
在单张 RTX 5090 上以 ISTA-DAS 2.25 bpw 量化本地跑 DeepSeek-V4.1-Flash：长生成 47 tok/s、prefill 1400 tok/s，同机胜过 Project Maya。
Lyes Sbahi
@lyesbahi12 小时前收录原推发布于 2026.10.28
Lyes Sbahi @lyesbahi原推发布于 2026.10.28
Day of tuning on Strata ⚡️→ then @coldniko shipped v0.2.7 and beat my numbers himself 😄 DeepSeek-V4.1-Flash (ISTA-DAS 2.25 bpw, RTX 5090): - 47 t/s long gen (2.2k tok bench) - 1,400 t/s prefill - beats Project Maya on the same box Qwen3.8-Flash-Next: 236 t/s, 6,672 t/s prefill
x.com/lyesbahi/status/2115159767312449818采集于 2026.10.28 12:03