Across 16 shared benchmarks, Nemotron 3 Super 120B A12B scores higher on 14 and NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16 on 2. The widest gap is RULER 1M, where Nemotron 3 Super 120B A12B scores 93.9 against 69.6.
| Benchmark | Nemotron 3 Super 120B A12B | NVIDIA-Nemotron-3.5-Lightning-30B-A3B-Base-BF16 |
|---|---|---|
| AGIEval-En | 77.9 | 70 |
| arc_challenge | 96.1 | 92.7 |
| Global-MMLU-Lite (5-shot) Average | 85.7 | 75.5 |
| GSM8K | 90.7 | 91.3 |
| hellaswag | 89 | 85.5 |
| humaneval | 80.5 | 77.4 |
| MBPP | 81.7 | 78.6 |
| Minerva Math | 84.8 | 82.8 |
| mmlu | 86 | 78.6 |
| MMLU-Pro | 83.8 | 67.9 |
| OpenBookQA | 48.6 | 47.6 |
| PIQA (acc-norm) | 85.5 | 85.2 |
| PIQA (acc) | 83.9 | 83.3 |
| RULER 1M | 93.9 | 69.6 |
| RULER 256K | 96.7 | 76.9 |
| winogrande | 78.9 | 80 |
Best tracked score per model per benchmark (default configuration; source-attributed). ↓ marks lower-is-better metrics. Open either model for its full surface, provenance and pricing.