NVIDIA-Nemotron-3-Super-120B-A12B-BF16 (Thinking): benchmark results
NVIDIA Nemotron 3 Super uses an 88-block hybrid Mamba-2/attention LatentMoE backbone: 40 Mamba, 40 MoE and eight attention blocks. Each MoE routes to 22 of 512 experts plus one shared expert; latent width is 1,024. A separate shared MTP module supports speculative decoding. Provider: NVIDIA. Released 2026-03-11. Access: Open.
Aggregate Rating 1609 ± 17, rank #691 of 2170 rated models, from 41 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Swallow - Post-trained English - AIME | 89.2 | Accuracy (%) | 93.3 |
| Swallow - Post-trained English - LiveCodeBench | 69.1 | Pass@1 (%) | 92.5 |
| Swallow - Post-trained English - IFBench | 66 | Instruction-Level Strict Accuracy (%) | 91 |
| Swallow - Post-trained English - Average | 82.5 | Average Score (%) | 89.6 |
| Swallow - Post-trained English - MATH-500 | 98.4 | Accuracy (%) | 89.6 |
| Swallow - Post-trained Japanese - M-IFEval-Ja | 82.3 | Instruction-Level Strict Accuracy (%) | 89.6 |
| Swallow - Post-trained English - GPQA Diamond | 81.8 | Accuracy (%) | 88.1 |
| Swallow - Post-trained Japanese - GPQA | 73.9 | Accuracy (%) | 87.3 |
| Swallow - Post-trained English - MMLU-Pro | 83.6 | Accuracy (%) | 86.6 |
| Swallow - Post-trained Japanese - Average | 61.1 | Average Score (%) | 86.6 |
| Swallow - English MT-Bench - Roleplay | 78.3 | Judge Score (normalized, %) | 85.1 |
| Swallow - Post-trained Japanese - PolyMath High and Top | 54.4 | Accuracy (%) | 85.1 |
Interactive version: theaggregate.ai/model?slug=nvidia-nemotron-3-super-120b-a12b-bf16-thinking · How It Works · Data refreshed daily, snapshot 2026-11-05.