Nemotron 3 Ultra (Reasoning): benchmark results
NVIDIA Nemotron 3 Ultra uses a 108-block hybrid Mamba-2/attention LatentMoE backbone: 48 Mamba, 48 MoE and 12 attention blocks. Each MoE routes to 22 of 512 experts plus one shared expert; latent width is 2,048. Shared-weight MTP prediction heads are separate from the backbone. Provider: NVIDIA. Released 2026-06-04. Access: Open.
Aggregate Rating 1605 ± 13, rank #743 of 2287 rated models, from 147 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| K2 Horizon MoVA Card - GPQA Diamond | 86.7 | Score (%) | 100 |
| K2 Horizon MoVA Card - Humanity's Last Exam (without tools) | 28.4 | Score (%) | 100 |
| SEA-HELM (English) - IFEval | 95.05 | Normalized Score | 100 |
| SEA-HELM (English) - MMLU Redux | 92.61 | Normalized Score | 98.3 |
| SEA-HELM (Filipino) - SEA-MT-Bench (LLM Judge) | 86.45 | Normalized Score | 97.3 |
| SEA-HELM (Indonesian) - Pragmatics | 79.5 | Normalized Score | 97.3 |
| SEA-HELM (Indonesian) - SEA-MT-Bench (LLM Judge) | 89.06 | Normalized Score | 97.3 |
| SEA-HELM (Vietnamese) - SEA-MT-Bench (LLM Judge) | 88.37 | Normalized Score | 97.3 |
| SEA-HELM (Malay) - SEA-MT-Bench (LLM Judge) | 87.47 | Normalized Score | 95.9 |
| SEA-HELM (Filipino) - Kalahi (LLM Judge) | 83.83 | Normalized Score | 94.6 |
| SEA-HELM (Filipino) - Toxicity Detection | 56.72 | Normalized Score | 94.6 |
| SEA-HELM (English) - MathArena | 47.27 | Normalized Score | 93.3 |
Interactive version: theaggregate.ai/model?slug=nemotron-3-ultra-reasoning · How It Works · Data refreshed daily, snapshot 2026-11-22.