Nemotron 3.5 Super
One DGX SparkOn one DGX Spark it decodes 46.0 tok/s and reads a 48,000-token codebase at 2,732 tok/s. It answers 83.6% of MMLU-Pro and gets 3 of 7 one-shot builds working. Ray tested it on 2026-10-20.
- Machine
- One DGX Spark: NVIDIA GB10, 128 GB unified memory
- Engine
- vLLM 0.32 nightly
- Weights
- NVFP4, official weights
- Recipe
- NVIDIA’s single-Spark recipe: LiLiCorr DFlash drafts (7), 262K context, one request at a time
- Made by
- NVIDIA
- Size
- 120B MoE, 12B active
- Tested
- 2026-10-20
- Asked who made it
- “I am a language model developed by researchers from NVIDIA.”
- Note
- Run by hand with NVIDIA’s staged launcher, outside Ownlight’s catalog. The accuracy sets ran on a second copy with four request slots and a 64K context, which changes speed, not answers.
- Overall
- 643
- Builds
- 643
- Knowledge
- 836
- Speed
- 449
Scores out of 1,000. Overall is the mean of the other three; how each is scored.
Speed
- Decode speed
- 46.0 tok/s
- Decode, prose
- 39.8 tok/s
- Decode, Snake
- 52.1 tok/s
- First token
- 363 ms
- Prefill, 48k prompt
- 2,732 tok/s
- Tokens for the seven builds
- 82,817
- Time for the seven builds
- 31m 25s
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 | 117 of 140 | 83.6% | 92.6 |
| GSM8K, 100 problems, thinking off | 97 of 100 | 97.0% | 1.5 |
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) | 39.8 | 363 ms | — | 2,002 | 50.7 s | 930 |
| Snake | 52.1 | 488 ms | — | 3,920 | 1m 15s | 932 |
| Read a 48k-token codebase | 43.8 | 16.3 s | 2,732 | 1,838 | 58.3 s | 933 |
| Who made you? | 54.3 | 303 ms | — | 72 | 1.6 s | 934 |
| Rocket launch | 43.4 | 408 ms | — | 10,696 | 4m 06s | 935 |
| Heptagon balls | 46.8 | 431 ms | — | 10,532 | 3m 45s | 936 |
| Solar system | 43.7 | 405 ms | — | 6,589 | 2m 31s | 937 |
| Pelican on a bicycle | 45.5 | 394 ms | — | 5,254 | 1m 55s | 938 |
| City fireworks | 42.4 | 413 ms | — | 14,491 | 5m 41s | 939 |
| Gummy bears | 43.1 | 536 ms | — | 31,335 | 12m 08s | 940 |
| Capybara onsen | 42.3 | 591 ms | — | 29,995 | 11m 50s | 941 |
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
- Qwen 3.8 Flash Next on one DGX Spark, 66.7 tok/s
- DeepSeek V4.1 Flash 1017 on two DGX Sparks, 78.6 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.