DeepSeek V4.1 Flash 748B on a single RTX 5090: Maya-S quant test
A 340GB Maya-S quant of the 748B model runs on one RTX 5090 with experts streamed from RAM and SSD; DeepSeek Harness built a voxel airship dock in 74.6 minutes at 22.9 tok/s code decode.
🚨 DeepSeek-V4.1-Flash-1017 on one RTX 5090 🚨 That's a 748B model on a single card. His Maya-S quant, 340 GB on disk: the card holds 1,970 of its 15,360 experts, the rest is read from 96 GB of RAM and the NVMe SSD. One slot. Agent: DeepSeek Harness 0.2.1-rc.2. One prompt, one try, no help. Thinking on, sampling from the DeepSeek model card: temperature 1.0, top_p 1.0. Airship dock: done in 74.6 min Tokens written: 98,217 Browser check: 9/10 Decode, code: 22.9 tok/s (separate speed run, thinking off) Decode, prose: 18.9 tok/s 500 words: about 35.1 s Eight requests at once: 19.7 tok/s in total Context: 131,072 tokens Steampunk Airship Dock, Brasshaven Station 7: a voxel tower with brass gears, a crane, a forge, airships and its own camera tour. Open: the context filled up at the end. The last request could only write 16 tokens, so the harness ended with an error. The last saved page counts, and it loads without errors. Answer limit in the build test: 262,144 tokens, the figure on the model card. The context is 131,072, so my test proxy lowered the limit of 24 of the 32 requests to what still fits. Measured on Project Maya v1.1.22. Newer releases have come out since. Play the world 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: Qwen3.8-27B on @MiaAI_lab's kit.