DeepSeek V4.1 Flash on one RTX 5090: 22.9 tok/s decode, 80% GPQA with thinking
The 340GB Maya-S quant keeps only 1,970 of 15,360 experts in VRAM, reading the rest from RAM and SSD; code decode hits 22.9 tok/s, first token after a 32K prompt takes 17.7s, MMLU-Pro is 70.8% and LiveCodeBench 60%.
🚨 DeepSeek-V4.1-Flash-1017 on one RTX 5090 🚨 Peasant Smith's own Maya-S quant, 340 GB on disk, in Project Maya v1.1.22 as its setup writes it: --tp 2, the 128K context option, one slot. Sampling from the DeepSeek model card and Maya's own default (temperature 1.0, top_p 1.0). Thinking off for every number here, except the GPQA line marked with thinking. The model does not fit on the card. Only 1,970 of its 15,360 experts sit in the 5090's 32 GB. The other 13,390 are read in place from 96 GB of RAM and the NVMe SSD, so most of the 340 GB is offloaded. Mia's Qwen3.8-27B in the same 32 GB class runs fully in VRAM, so the two are not built the same way. Decode, code: 22.9 tok/s Decode, prose: 18.9 tok/s 500 words: about 35.1 s Reading a 32K-token prompt: 1,847 tok/s First token after a 32K prompt: 17.7 s Two requests at once: 18.9 tok/s in total Four requests at once: 19.5 tok/s in total Eight requests at once: 19.7 tok/s in total Context: 131,072 tokens MMLU-Pro: 70.8% ± 4.4 points GPQA Diamond: 52.5% ± 6.9 points IFEval: 88.0% ± 4.5 points LiveCodeBench: 60.0% ± 14.5 points Long text 8K: 72.5% ± 13.4 points Long text 32K: 52.5% ± 14.8 points GPQA with thinking: 80.0% ± 13.9 points The server has one slot, so parallel requests take turns. Open: MMLU-Pro and GPQA answers were capped at 1,024 tokens. 2 of 400 and 9 of 198 hit the cap and count as wrong. 1 long-text answer and 5 of 30 GPQA answers with thinking also ran past our token limits. Open: a single request measured end to end swung between about 8 and 17 tok/s from round to round (the wait for the first token and the speed both varied). It is not shown. Measured on Project Maya v1.1.22. Newer releases have come out since. Play the airship dock it built in your browser. Every number, the exact command and the raw output: https://niklaslenz.com/v2/deepseek-v41-flash-5090-maya-s-mayads4/ Project Maya and the Maya-S quant by @PeasantSmith, co-maintained by needmorevram. DeepSeek-V4.1-Flash-1017 by @deepseek_ai. My pick on 32 GB: Still Qwen3.8-27B on @MiaAI_lab's kit. It is just WAY faster, has almost the same quality and does not need the RAM.