Nanbeige4-3B-Thinking-2511: benchmark results
Chinese Nanbeige Lab's 3B thinking model (November 2025 update), distillation- and RL-trained for outsized function-calling and writing performance. Provider: Nanbeige. Released 2025-11-20. Access: Open.
Aggregate Rating 1476 ± 25, rank #2720 of 10916 rated models, from 16 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| WritingBench | 79.03 | Score (imported) | 79.6 |
| UGI - Willingness (W/10) | 4.5 | W/10 Score | 35.6 |
| JP-TL-Bench - Japanese to English - Easy | 4.29 | LT Score (0-10; Easy items only; anchored pairwise Bradley-T | 27 |
| JP-TL-Bench - Japanese to English | 2.87 | LT Score (0-10; anchored pairwise Bradley-Terry vs 20 Base S | 22.2 |
| JP-TL-Bench - Japanese to English - Hard | 1.93 | LT Score (0-10; Hard items only; anchored pairwise Bradley-T | 22.2 |
| UGI Leaderboard | 24.14 | UGI Score | 17.6 |
| EQ-Bench Creative Writing v3 | 814.6 | Elo | 17.3 |
| Creative Writing v3 | 965.5 | Elo score (imported) | 16.4 |
| JP-TL-Bench - English to Japanese | 0.87 | LT Score (0-10; anchored pairwise Bradley-Terry vs 20 Base S | 15.9 |
| JP-TL-Bench - English to Japanese - Hard | 1.04 | LT Score (0-10; Hard items only; anchored pairwise Bradley-T | 15.9 |
| UGI - Natural Intelligence | 13.72 | NatInt Score | 13.2 |
| JP-TL-Bench - English to Japanese - Easy | 0.58 | LT Score (0-10; Easy items only; anchored pairwise Bradley-T | 9.5 |
Interactive version: theaggregate.ai/model?slug=nanbeige4-3b-thinking-2511 · How It Works · Data refreshed daily, snapshot 2026-11-08.