Nemotron 3 Super: benchmark results
NVIDIA's open 120B-total/12B-active hybrid Mamba-Transformer MoE (March 2026) with 1M context, tuned for high-throughput agentic inference. Provider: NVIDIA. Released 2026-03-11. Access: Open.
Aggregate Rating 1638 ± 9, rank #335 of 1656 rated models, from 385 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| ACTS - Blind Cipher Identification | 37.3 | Accuracy (%; tier-3 blind identification of the algorithm fa | 100 |
| LLM Stats (WMT24++) | 86.67 | Score (%) | 100 |
| NVIDIA Nemotron 3.5 Lightning Card - AA-Omniscience | 26.68 | Accuracy (%) | 100 |
| NVIDIA Nemotron 3.5 Lightning Card - HLE (text-only, no tools) | 20.3 | Score (%) | 100 |
| PPT-Bench (Epistemic Attack) | 23.3 | Binary capitulation rate (%): share of items scored 1 or 2 o | 100 |
| PPT-Bench (Epistemic Attack) - Identity Dissolution | 17.4 | Binary capitulation rate (%): share of items scored 1 or 2 o | 100 |
| PPT-Bench (Epistemic Attack) - Value Nullification | 9.1 | Binary capitulation rate (%): share of items scored 1 or 2 o | 100 |
| Ukrainian Legal Text Benchmark (EDRSR subset) | 83.1 | 3-task Composite (self-reported) | 100 |
| UnpredictaBench - Code | 40.34 | KS@100 (%): share of the 238 Code problems (distributions in | 100 |
| AGC-Bench - grapheval_ai_researcher | 1.42 | Dataset z-score | 98.8 |
| UnpredictaBench - Text | 28.13 | KS@100 (%): share of the 160 Text problems (named or describ | 97.4 |
| FinRiskAtlas - Applied Review - Disputed-Issue Generation | 76.67 | Rubric Score (%; DeepSeek-V4-Flash judge) | 96.8 |
Interactive version: theaggregate.ai/model?slug=nemotron-3-super · How It Works · Data refreshed daily, snapshot 2026-10-20.