P3B3 (pt-BR Prompt): leaderboard
P3B3 (2026) measures European versus Brazilian Portuguese variety bias and controllability in expert-written, variety-agnostic multi-turn dialogues. This board scores the replies when the first turn asks for Brazilian Portuguese, on a scale from 0 (Brazilian) to 100 (European Portuguese) judged by Gemini 3 Flash; lower is better, meaning the model follows the requested variety.
Metric: Variety score (0-100), where 0 is Brazilian and 100 European Portuguese, of the model replies when the first turn asks for Brazilian Portuguese; Gemini-3-Flash judge, Portuguese category-based single-turn prompt, averaged over 203 turns; greedy decoding; lower is better (follows the requested variety). Source: arxiv.org. Saturation forecast: Estimated already saturated. 20 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Preview) | 4.1 | #144 |
| 2 | gemma-4-E4B-it | 4.2 | #1469 |
| 3 | Ministral-3-14B-Instruct-2512 | 4.2 | #1593 |
| 4 | Gemma 3 12B (IT) | 4.9 | #1837 |
| 5 | Llama 3.3 70B Instruct | 6.7 | #1357 |
| 6 | Gemma 4 31B (IT) | 7.3 | #342 |
| 7 | Qwen 3 8B | 7.9 | #1830 |
| 8 | Apertus-70B-Instruct-2509 | 8.6 | #2639 |
| 9 | EuroLLM-22B-Instruct-2512 | 9.6 | #1994 |
| 10 | OLMo 3.1 32B Instruct | 10.5 | #2384 |
| 11 | Llama 3.1 8B Instruct | 10.6 | #4246 |
| 12 | Qwen 3.5 27B | 11.9 | #359 |
| 13 | Qwen 3.5 9B | 12.2 | #803 |
| 14 | Apertus-8B-Instruct-2509 | 14.7 | #3960 |
| 15 | OLMo 3 7B Instruct | 15.8 | #6860 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=p3b3-pt-br-prompt · How It Works · Data refreshed daily, snapshot 2026-11-13.