deepseek-v4.1IMAGE
DeepSeek V4.1 Flash at 2.25bpw on RTX 5090: ~50 tok/s generation, 1350 tok/s prefill
Tuning Strata on an RTX 5090 with a 2.25bpw 32GB quantization of V4.1 Flash yields ~50–51 tok/s long generation at 256K context and 1350 tok/s cold prefill.
Lyes Sbahi
@lyesbahiArchived 2 days agoOriginally posted 2026.10.27
Lyes Sbahi @lyesbahiOriginally posted 2026.10.27
Day of tuning on Strata ⚡️ No magic "Turbo" switch yet, but solid raw metrics on DeepSeek-V4.1-Flash (2.25 bpw / 32 GB) using the RTX 5090: 🔹 ~50-51 t/s long gen avg (2.2k tokens bench, 256K ctx) 🔹 1,350 t/s prefills @coldniko
x.com/lyesbahi/status/2114858533984629107Archived 2026.10.27 12:08