DeepSeek V4.1 Flash 1017
Two DGX SparksOn two DGX Sparks it decodes 78.6 tok/s and reads a 48,000-token codebase at 1,919 tok/s. It answers 87.9% of MMLU-Pro and gets 7 of 7 one-shot builds working. Ray tested it on 2026-10-17.
- Machine
- Two DGX Sparks: 2 × NVIDIA GB10, 2 × 128 GB unified memory, linked at 200 Gb/s
- Engine
- TensorFold Zig (Bertholomus AI)
- Weights
- EXL3 2.9 bpw community quant (LibertAIDAI)
- Recipe
- Tensor parallel over both Sparks, 1M context
- Made by
- DeepSeek
- Size
- MoE
- Tested
- 2026-10-17
- Asked who made it
- “I was made by DeepSeek.”
- Note
- The gummy-bear page is run 822 (2026-10-18), the fair rerun at the model’s own sampling with the 262,144-token budget; the first suite run hit a 65,536 cap.
- Overall
- 828
- Builds
- 1,000
- Knowledge
- 879
- Speed
- 606
Scores out of 1,000. Overall is the mean of the other three; how each is scored.
Speed
- Decode speed
- 78.6 tok/s
- Decode, prose
- 60.6 tok/s
- Decode, Snake
- 96.5 tok/s
- First token
- 365 ms
- Prefill, 48k prompt
- 1,919 tok/s
- Tokens for the seven builds
- 204,621
- Time for the seven builds
- 42m 43s
Decode speed is the median of a prose prompt and a Snake game, one request at a time. Prefill is how fast it reads a 48,000-token codebase before it answers.
Accuracy
| Set | Correct | Score | Minutes |
|---|---|---|---|
| MMLU-Pro, 140 questions, thinking low | 123 of 140 | 87.9% | 11.7 |
| GSM8K, 100 problems, thinking low | 99 of 100 | 99.0% | 2.5 |
Builds
Run these pagesRun local models in Ownlight
Ownlight connects the model servers you already run, like Ollama, LM Studio, Unsloth Studio, llama-swap or Model Deck, and starts Claude Code or Codex on a local model beside your cloud agents. Local is part of Free.
Every step
Decode speed, time to first token, prefill speed, output tokens and total time for each prompt in the suite.
| Test | Decode tok/s | First token | Prefill tok/s | Output tokens | Total time | Run |
|---|---|---|---|---|---|---|
| Prose (speed) | 60.6 | 365 ms | — | 1,786 | 29.8 s | 798 |
| Snake | 96.5 | 662 ms | — | 6,967 | 1m 12s | 800 |
| Read a 48k-token codebase | 76.3 | 23.7 s | 1,919 | 933 | 35.9 s | 801 |
| Who made you? | 80.7 | 256 ms | — | 51 | 890 ms | 802 |
| Rocket launch | 81.0 | 424 ms | — | 51,372 | 10m 35s | 805 |
| Heptagon balls | 74.7 | 416 ms | — | 21,800 | 4m 52s | 806 |
| Solar system | 82.3 | 2.1 s | — | 20,284 | 4m 08s | 807 |
| Pelican on a bicycle | 81.8 | 362 ms | — | 24,441 | 4m 59s | 808 |
| City fireworks | 79.3 | 461 ms | — | 39,096 | 8m 13s | 809 |
| Gummy bears | 77.9 | 583 ms | — | 40,661 | 8m 42s | 822 |
| Capybara onsen | 79.4 | 1.0 s | — | 73,183 | 15m 22s | 915 |
More models on the bench
- Qwen 3.8 27B on one RTX 5090, 189.2 tok/s
- Qwen 3.8 27B on one DGX Spark, 31.8 tok/s
- Nemotron 3.5 Super on one DGX Spark, 46.0 tok/s
- Qwen 3.8 Flash Next on one DGX Spark, 66.7 tok/s
- GLM 5.3 Flash (uncensored) on two DGX Sparks, 55.6 tok/s
- DeepSeek V4.1 Flash (September) on two DGX Sparks, 81.9 tok/s, retired
Updated 2026-10-20 16:39.