NVIDIA AI & NIMopen-weight
Released: 2025-04-10Nemotron 3 Ultra 550B
NVIDIA 550B MoE workhorse ranking among the top models on OpenRouter, delivering high-speed inference across 8x H100 clusters.
Intelligence Index
91.2/ 100
Calibrated multi-domain compositeThroughput Speed
85tok/sec
Output streaming throughputTime To First Token
340ms
Initial chunk server latencyLiveBench Contam-Free
78.5%
September 2026 suite scoreSourced Benchmark Evaluations (61)
Verified performance across authoritative benchmarks with provenance tracking.| Benchmark | Status | Score | Trust | Date | Source Type | Provenance |
|---|---|---|---|---|---|---|
| ARC-AGI-3 | Active | 59.2% | 93 | 2025-05-15 | independent | Source ↗ |
| ARC-AGI-2 | Nearing Saturation | 71% | 75 | 2025-05-08 | independent | Source ↗ |
| FrontierMath | Active | 25.3% | 98 | 2025-05-05 | independent | Source ↗ |
| PutnamBench | Active | 47.1% | 90 | 2025-05-04 | independent | Source ↗ |
| SWE-Lancer | Deprecated | 56.8% | 20 | 2025-05-02 | independent | Source ↗ |
| Omni-MATH | Deprecated | 67% | 10 | 2025-05-02 | independent | Source ↗ |
| Soohak | Active | 68.8% | 98 | 2025-04-30 | independent | Source ↗ |
| FrontierCode | Active | 48.4% | 98 | 2025-04-30 | independent | Source ↗ |
| Humanity's Last Exam | Active | 40% | 93 | 2025-04-28 | independent | Source ↗ |
| StructFlowBench | Nearing Saturation | 71.3% | 75 | 2025-04-28 | independent | Source ↗ |
| AIME | Nearing Saturation | 86% | 75 | 2025-04-28 | independent | Source ↗ |
| OlympiadBench | Active | 74.2% | 80 | 2025-04-27 | independent | Source ↗ |
| MathArena | Active | 72.4% | 99 | 2025-04-26 | independent | Source ↗ |
| FireBench | Active | 70.5% | 95 | 2025-04-26 | independent | Source ↗ |
| Terminal-Bench | Active | 58.5% | 80 | 2025-04-26 | independent | Source ↗ |
| OpenAI-MRCR | Active | 76.6% | 80 | 2025-04-25 | independent | Source ↗ |
| SimpleQA | Active | 47.4% | 90 | 2025-04-25 | independent | Source ↗ |
| CodeContests | Deprecated | 54.3% | 10 | 2025-04-25 | independent | Source ↗ |
| ComplexBench | Nearing Saturation | 68.8% | 80 | 2025-04-24 | independent | Source ↗ |
| SWE-bench Pro | Active | 51.8% | 90 | 2025-04-24 | independent | Source ↗ |
| LongBench-Pro | Active | 74.8% | 80 | 2025-04-24 | independent | Source ↗ |
| ZeroSCROLLS | Nearing Saturation | 73% | 78 | 2025-04-23 | independent | Source ↗ |
| MultiPL-E | Deprecated | 76.8% | 10 | 2025-04-23 | independent | Source ↗ |
| LiveCodeBench | Saturated | 68.5% | 39 | 2025-04-22 | independent | Source ↗ |
| AgentIF | Nearing Saturation | 57.1% | 80 | 2025-04-22 | independent | Source ↗ |
| MATH | Saturated | 88.7% | 25 | 2025-04-22 | independent | Source ↗ |
| InfiniteBench | Nearing Saturation | 77.5% | 65 | 2025-04-22 | independent | Source ↗ |
| Mercury | Deprecated | 73.5% | 10 | 2025-04-21 | independent | Source ↗ |
| SCROLLS | Nearing Saturation | 73% | 65 | 2025-04-21 | independent | Source ↗ |
| MMLU-Pro | Nearing Saturation | 83.9% | 75 | 2025-04-21 | independent | Source ↗ |
| RULER | Nearing Saturation | 83.9% | 84 | 2025-04-20 | independent | Source ↗ |
| TabMWP | Saturated | 83.3% | 25 | 2025-04-20 | independent | Source ↗ |
| CyberSecEval | Deprecated | 75.2% | 10 | 2025-04-20 | independent | Source ↗ |
| SWE-bench Verified | Saturated | 65.1% | 25 | 2025-04-19 | independent | Source ↗ |
| MMLU-Redux | Saturated | 89.4% | 25 | 2025-04-19 | independent | Source ↗ |
| QuALITY | Nearing Saturation | 80.3% | 78 | 2025-04-19 | independent | Source ↗ |
| LongBench | Nearing Saturation | 78.4% | 65 | 2025-04-18 | independent | Source ↗ |
| SWE-bench | Deprecated | 56.8% | 10 | 2025-04-18 | independent | Source ↗ |
| MGSM | Nearing Saturation | 86% | 65 | 2025-04-18 | independent | Source ↗ |
| AGIEval | Saturated | 78.3% | 25 | 2025-04-17 | independent | Source ↗ |
| QASPER | Nearing Saturation | 75.7% | 65 | 2025-04-17 | independent | Source ↗ |
| TruthfulQA | Saturated | 80.3% | 25 | 2025-04-17 | independent | Source ↗ |
| GSM-Hard | Deprecated | 86.9% | 10 | 2025-04-16 | independent | Source ↗ |
| SuperGLUE | Saturated | 92.8% | 25 | 2025-04-16 | independent | Source ↗ |
| CodeXGLUE | Deprecated | 78.5% | 23 | 2025-04-15 | independent | Source ↗ |
| MT-Bench | Nearing Saturation | 8.6% | 75 | 2025-04-15 | independent | Source ↗ |
| BIG-Bench Hard | Saturated | 89.2% | 25 | 2025-04-15 | independent | Source ↗ |
| WinoGrande | Saturated | 94.6% | 28 | 2025-04-14 | independent | Source ↗ |
| MBPP+ | Nearing Saturation | 85.2% | 65 | 2025-04-14 | independent | Source ↗ |
| IFEval | Deprecated | 82.2% | 25 | 2025-04-14 | independent | Source ↗ |
| GPQA | Nearing Saturation | 80.3% | 78 | 2025-04-14 | independent | Source ↗ |
| Needle-in-a-Haystack | Saturated | 98.2% | 44 | 2025-04-13 | independent | Source ↗ |
| HumanEval+ | Saturated | 90.2% | 25 | 2025-04-13 | independent | Source ↗ |
| CommonsenseQA | Saturated | 95.6% | 28 | 2025-04-13 | independent | Source ↗ |
| MBPP | Saturated | 88.5% | 25 | 2025-04-12 | independent | Source ↗ |
| GSM8K | Saturated | 97.7% | 25 | 2025-04-12 | independent | Source ↗ |
| GLUE | Saturated | 98.3% | 28 | 2025-04-12 | independent | Source ↗ |
| MMLU | Saturated | 93% | 25 | 2025-04-11 | independent | Source ↗ |
| HumanEval | Saturated | 95.2% | 25 | 2025-04-11 | independent | Source ↗ |
| ARC-AGI | Saturated | 85.5% | 25 | 2025-04-11 | independent | Source ↗ |
| LiveBench (Continuous) | Active | 78.5% | 94 | 2025-02-01 | independent | Source ↗ |
Local Hardware Execution (Ollama)Hardware Compatible
# 1. Pull and execute model locally with Ollama
ollama run nvidia-nemotron-3-ultra-550b
# 2. Or invoke via local OpenAI-compatible endpoint
curl http://localhost:11434/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "nvidia-nemotron-3-ultra-550b",
"messages": [{"role": "user", "content": "Analyze reasoning chains on AIME 2026"}]
}'Evaluation & Sourcing Notes
Scores listed for Nemotron 3 Ultra 550B represent verified evaluations extracted from official research papers, independent evaluation suites (HELM, LMSYS, OpenCompass, LiveBench), and verified audit reports.
All benchmarks marked as saturated or deprecated reflect historical performance where the benchmark no longer provides active discriminative power.