Capabilities
- MMLU 89.1
- MMLU-Pro 79.1
- AGIEval 78.7
- GPQA Diamond 50
- GSM8K 88.1
- MATH 82
- HumanEval 83.8
- MBPP 86
- ARC-Challenge 97.3
- HellaSwag 90.5
- OpenBookQA 48.6
- PIQA 83.8
- WinoGrande 79.3
- RACE 92.2
- Global MMLU Lite 3.0 90.1
- MGSM 87.7
- RULER 92.5
Nemotron 3 Ultra 550B-A55B Base
Nemotron 3 Ultra 550B-A55B Base ranks first on ARC-Challenge.
17
Benchmarks with a score
1
Ranked first, out of five or more models
2
In the top three
NVIDIA Jun 2026 Nemotron 3 Ultra 550B-A55B Base
Open weights Flagship
127.3
#296 of 404
Self-Reported Index, from the lab's own numbers. No Epoch Capabilities Index score yet.
- MMLU-Pro 79.1% #86 /167
- MMLU 89.1% #21 /131
- RULER 92.5% #14 /42
Where it ranks
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
Knowledge and language
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| ARC-Challenge | 97.3% | #1 of 49 | Lab model card | |
| HellaSwag | 90.5% | #7 of 53 | Lab model card | |
| MMLU | 89.1% | #21 of 131 | Lab model card | |
| AGIEval | 78.7% | #8 of 40 | Lab model card | |
| RACE | 92.2% | #2 of 8 | Lab model card | |
| Global MMLU Lite 3.0 | 90.1% | #1 of 4 | Lab model card | |
| MGSM | 87.7% | #17 of 57 | Lab model card | |
| RULER | 92.5% | #14 of 42 | Lab model card | |
| PIQA | 83.8% | #14 of 37 | Lab model card | |
| WinoGrande | 79.3% | #22 of 47 | Lab model card | |
| MMLU-Pro | 79.1% | #86 of 167 | Lab model card | |
| OpenBookQA | 48.6% | #18 of 24 | Lab model card |
Math and science
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| MATH | 82% | #22 of 99 | Lab model card | |
| GSM8K | 88.1% | #50 of 92 | Lab model card | |
| GPQA Diamond | 50% | #246 of 302 | Lab model card |
Its lab's numbers put it at 127.3, number 296 of 404
The Self-Reported Index fits one score to the 16 capability results NVIDIA reported for this model, on the scale of the Epoch Capabilities Index. Labs choose which benchmarks to show and how to run them, so read it as the lab's claim, not an independent measurement. The 90% interval runs from 123.5 to 131.2. How we compute it.
- ARC-Challenge 97.3
- RULER 92.5
- RACE 92.2
- HellaSwag 90.5
- MMLU 89.1
- GSM8K 88.1
- MGSM 87.7
- MBPP 86
- HumanEval 83.8
- PIQA 83.8
- MATH 82
- WinoGrande 79.3
- MMLU-Pro 79.1
- AGIEval 78.7
- GPQA Diamond 50
- OpenBookQA 48.6
Its system card
Every result in the longer documents the lab published with it. 17 of them aren't in the launch post.