Models & Hardware
DeepSeek-1017 on One RTX 5090: 22.9 Tok/s From a 340 GB Quant
The card holds 1,970 of 15,360 experts. The rest stream from RAM and an SSD.
A 748B model on one consumer card, mostly offloaded.
@niklaslenz_ai ran DeepSeek-V4.1-Flash-1017 on a single RTX 5090 with @PeasantSmith’s Maya-S quant in Project Maya v1.1.22. The quant is 340 GB on disk. The card’s 32 GB hold 1,970 of the 15,360 experts; the other 13,390 are read in place from 96 GB of RAM and the NVMe SSD.
| Test | Result |
|---|---|
| Decode, code | 22.9 tok/s |
| Decode, prose | 18.9 tok/s |
| Reading a 32K prompt | 1,847 tok/s |
| Eight requests at once | 19.7 tok/s in total |
| MMLU-Pro | 70.8% |
| IFEval | 88.0% |
In his build test, the model wrote a steampunk airship dock in 74.6 minutes and scored 9 of 10, though the context filled up at the end. For comparison, his two-Spark run of @sudoingX’s recipe measured 64.9 tok/s on code. His pick on 32 GB remains Qwen3.8-27B on @MiaAI_lab’s kit, which runs fully in VRAM.
Sources
- @niklaslenz_ai on X · Oct 23, 11:25 AM PT
- @niklaslenz_ai on X · Oct 23, 11:00 AM PT