Models · Compare
Llama-3.1-Nemotron Ultra 253B v1 vs OpenAI o3
Rows highlighted in warm gray are where the models differ. Numbers carry their as-of date and primary source.
Specs
| Field | A: Llama-3.1-Nemotron Ultra 253B v1 | B: OpenAI o3 |
|---|---|---|
| Released | 2025-04-11 | 2025-04-16 |
| Developer | NVIDIA | OpenAI |
| Openness | Open weights | Proprietary |
| License | NVIDIA Open Model License | Proprietary |
| OSI-approved | no | no |
| Data released | yes | no |
| Training code | no | no |
| Architecture | dense | unknown |
| Total params | 253B | — |
| Active params | — | — |
| Experts | — | — |
| Context window | 131K | 200K |
| Attention | skip-attention | unknown |
| Position enc. | rope | unknown |
| Pretraining tokens | 65B | — |
| Post-training | sft, grpo | rlhf |
| Training hardware | H100 | — |
| $/M input | — | $2.00 |
| $/M output | — | $8.00 |
| Output tok/sec | — | 88 |
Benchmarks
Missing scores render as not reported; never inferred. Bold highlights the leader per benchmark.
General reasoning
| MMLU-Pro | — | 85.3 2026-05-21 |
| GPQA-Diamond | 76.0 2025-04-11 | 87.7 2025-04-16 |
Code
| SWE-Bench Verified | — | 71.7 2025-04-16 |
| LiveCodeBench | 66.3 2025-04-11 | 80.8 2026-05-21 |
Math
| MATH | 97.0 2025-04-11 | 99.2 2026-05-21 |
| AIME 2024 | — | 90.3 2026-05-21 |
| AIME 2025 | 72.5 2025-04-11 | 88.3 2026-05-21 |
Held-out / arena
| IFEval | 88.8 2025-04-11 | — |
Context · A
Top of NVIDIA's Llama-Nemotron family, distilled from Llama 3.1 405B via Neural Architecture Search with skip attention, variable FFN, and FFN fusion. Released April 11 2025; single-node 8x H100 BF16 inference, 4x H100 FP8. Post-trained through SFT and GRPO RL stages.
Context · B
Released April 16, 2025 as the full-size successor to o1 in the o-series reasoning lineage, with multimodal (text plus image) input and a 200K-token context. OpenAI reported large jumps over o1 on SWE-bench Verified and Codeforces and roughly 3x the accuracy of o1 on ARC-AGI.
Llama-3.1-Nemotron Ultra 253B v1 detail → · OpenAI o3 detail →