Nemotron 3 Nano 30B: benchmark results
NVIDIA Nemotron 3 Nano 30B-A3B uses 23 Mamba-2 blocks, 23 MoE blocks and six attention blocks. Each MoE routes to six of 128 experts plus one shared expert; base, post-trained reasoning modes and checkpoint precision variants remain distinct. Provider: NVIDIA. Released 2025-12-15. Access: Open.
Aggregate Rating 1573 ± 31, rank #1354 of 10916 rated models, from 13 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| SeekerGym - ML Surveys | 28.9 | Information-gathering completeness (%): share of a survey pa | 80 |
| ClaimDB: A Fact Verification Benchmark over Large Structured Data | 0.67 | Accuracy | 72.4 |
| SeekerGym - Wikipedia | 40.5 | Information-gathering completeness (%): share of a document' | 60 |
| Arcee MMLU-Pro (BenchLM) | 78.3 | Accuracy (%) | 44.6 |
| LEDGER - Comprehensive KPI Extraction | 65.5 | Recall (%) of the 13,519 labelled KPI cells of the 494 test | 33.3 |
| Structured Output Benchmark | 84.1 | Overall (%) | 31 |
| Wolfram LLM Benchmarking Project | 27.6 | Correct Functionality (%) | 25.9 |
| Sonar LLM Leaderboard - Java - Vulnerability Density | 0.38 | Vulnerabilities per 1,000 lines of code (lower is better) | 20 |
| SEA-HELM | 46.89 | Mean Score (%) | 19.3 |
| Sonar LLM Leaderboard - Java - Pass Rate | 71.3 | Passing tests (%) | 16 |
| Sonar LLM Leaderboard - Java - Bug Density | 1.16 | Bugs per 1,000 lines of code (lower is better) | 13.3 |
| Sonar LLM Leaderboard - Java - Code Smell Density | 22.57 | Code smells per 1,000 lines of code (lower is better) | 12 |
Interactive version: theaggregate.ai/model?slug=nemotron-3-nano-30b · How It Works · Data refreshed daily, snapshot 2026-11-08.