Model Details
Nemotron 3 Ultra 550B A55B
Quick answer: NVIDIA Nemotron 3 Ultra 550B A55B is NVIDIA's June 2025 MoE flagship scoring 87.0% GPQA Diamond, 94.2% AIME 2026, 37.4% HLE, 44.4% BrowseComp, 44.6% SciCode, and 86.8% MMLU-Pro. NVIDIA Open Model License with 550B total / 55B active parameters.
At a Glance
Where Nemotron 3 Ultra 550B leads
- 87.0% GPQA Diamond — strong graduate-level science
- 94.2% AIME 2026 — excellent competition mathematics
- 81.4% IFBench — strong instruction following
- 44.6% SciCode — competitive scientific coding
- NVIDIA Open Model License — available for research and production
Where it lags
- 37.4% HLE — below Kimi K2.6 (36.4%) and frontier models
- 44.4% BrowseComp — below Kimi K2.6 (83.2%) and GPT-5.5 (84.4%)
- Very large: 550B params, 55B active — requires significant infrastructure
Best for: NVIDIA NIM deployments; enterprise customers on NVIDIA AI infrastructure; large-scale scientific reasoning tasks.
What Nemotron 3 Ultra 550B A55B Is
Nemotron 3 Ultra 550B A55B is NVIDIA's June 2025 flagship MoE model — the largest in the Nemotron 3 generation. As a successor to Llama-based Nemotron variants (like Nemotron Ultra 253B), this model is NVIDIA's own MoE architecture with 550B total parameters and 55B active per token.
At the GPQA Diamond (87%) and AIME (94.2%) tier, it competes with strong frontier models for STEM reasoning, though it is below Kimi K2.6 and GLM-5.2 on HLE and BrowseComp tasks.
Specifications
| Field | Value |
|---|---|
| Organization | NVIDIA |
| License | NVIDIA Open Model License |
| HuggingFace | nvidia/Nemotron-3-Ultra-550B-A55B |
| Release date | June 2025 |
| Parameters | 550B total / 55B active (MoE) |
| Modality | Text only |
Pricing
Available under NVIDIA Open Model License — check for usage restrictions. Available via NVIDIA NIM API and cloud providers.
Public Benchmark Scores
| Benchmark | Score | Source | Date |
|---|---|---|---|
| GPQA Diamond | 87.0% | Benchgen evaluation | 2025-06 |
| AIME 2026 | 94.2% | Benchgen evaluation | 2025-06 |
| Humanity's Last Exam | 37.4% | Benchgen evaluation | 2025-06 |
| BrowseComp | 44.4% | Benchgen evaluation | 2025-06 |
| SimpleQA | 32.4% | Benchgen evaluation | 2025-06 |
| MMLU-Pro | 86.8% | Benchgen evaluation | 2025-06 |
| SciCode | 44.6% | Benchgen evaluation | 2025-06 |
| IFBench | 81.4% | Benchgen evaluation | 2025-06 |
Nemotron 3 Ultra 550B vs Alternatives
| Model | GPQA Diamond | HLE | MMLU-Pro | License |
|---|---|---|---|---|
| Nemotron 3 Ultra 550B | 87.0% | 37.4% | 86.8% | NVIDIA Open |
| Kimi K2.6 | 90.5% | 36.4% | — | Apache 2.0 |
| Llama 3.1 Nemotron Ultra 253B | — | — | — | Llama 3.1 |
| DeepSeek-R1 0528 | 81.0% | — | 85% | MIT |
Nemotron 3 Ultra 550B vs Kimi K2.6: similar HLE tier (37.4% vs 36.4%), lower GPQA Diamond (87% vs 90.5%), worse BrowseComp (44.4% vs 83.2%). Kimi K2.6 is superior on most metrics with Apache 2.0 license.
Related Models & Benchmarks
- Compare: Kimi K2.6 · Llama 3.1 Nemotron Ultra 253B · DeepSeek-R1 0528
- Benchmarks: GPQA Diamond · AIME 2026 · BrowseComp
Frequently Asked Questions
What is Nemotron 3 Ultra 550B A55B?
NVIDIA's June 2025 MoE flagship: 550B total / 55B active parameters, scoring 87.0% GPQA Diamond, 94.2% AIME, 37.4% HLE, 86.8% MMLU-Pro. NVIDIA Open Model License.How does Nemotron 3 Ultra 550B compare to Kimi K2.6?
Kimi K2.6 outperforms on most metrics: higher GPQA Diamond (90.5% vs 87%), much higher BrowseComp (83.2% vs 44.4%), better SWE-Bench (80.2%). Kimi K2.6 is Apache 2.0. Nemotron 3 Ultra suits teams within the NVIDIA NIM ecosystem.Specs from NVIDIA's Nemotron 3 Ultra 550B A55B release (June 2025) and Benchgen evaluations. Last updated 2026-07-24.
Benchmark Leaderboards
This model isn’t on any benchmark leaderboard yet.