Nanbeige4.1-3B: benchmark results
Nanbeige Lab's 3B reasoning and agentic model, an SFT+RL post-train of its Nanbeige4-3B base aimed at matching much larger Qwen3 models (February 2026). Provider: Nanbeige. Released 2026-02-11. Access: Open.
Unified ELO 1464 ± 1, rank #879 of 1392 rated models, from 37 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| AA GPQA Diamond | 84.95 | Accuracy (%) | 78 |
| OpenEvals - GPQA Diamond | 83.8 | Accuracy (%) | 65.4 |
| AA Humanity's Last Exam | 10.89 | Accuracy (%) | 56.7 |
| AA Omniscience | -42.5 | Score | 45.2 |
| Artificial Analysis Intelligence Index | 5.34 | Intelligence Index | 35.7 |
| AA Omniscience - Software Engineering (SWE) - Julia | 4 | Accuracy (%) | 30.4 |
| BenchmarkList ECI | 96.02 | Capability Index (ECI) | 28.9 |
| AA IFBench | 35.44 | Accuracy (%) | 27.4 |
| AA CritPt | 0 | Accuracy (%) | 24.4 |
| UGI - Willingness (W/10) | 3 | W/10 Score | 22.7 |
| AA TAU-2 Bench | 21.64 | Accuracy (%) | 22.4 |
| AA Omniscience - Science, Engineering & Mathematics | 19.2 | Accuracy (%) | 21.9 |
Interactive version: theaggregate.ai/model?slug=nanbeige4-1-3b · How It Works · Data refreshed daily, snapshot 2026-09-05.