MT-Bench PL - Reasoning: leaderboard
The "Reasoning" subtask of MT-Bench PL. The Polish-language version of MT-Bench, scoring multi-turn conversational and instruction-following ability via LLM-as-judge.
Metric: Judge Score (0-10). Source: huggingface.co. Saturation forecast: Estimated already saturated. 50 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Phi-4 | 9.55 | #706 |
| 2 | Qwen 2.5 32B Instruct | 9.1 | #498 |
| 3 | Mistral Small 3.1 | 9 | #610 |
| 4 | aya-expanse-32B | 8.95 | #692 |
| 5 | Qwen 2 72B Instruct | 8.85 | #576 |
| 6 | Mistral Large 2 (Jul) | 8.7 | #484 |
| 7 | Gemma 3 27B (IT) | 8.4 | #523 |
| 8 | Bielik-11B-v2.3-Instruct | 8.35 | #693 |
| 9 | Mistral Small 3 | 7.9 | #627 |
| 10 | Mistral-Small-Instruct-2409 | 7.9 | #702 |
| 11 | Gemma 3 12B (IT) | 7.75 | #664 |
| 12 | Qwen 2.5 14B Instruct | 7.55 | #647 |
| 13 | Bielik-11B-v2.2-Instruct | 6.9 | #660 |
| 14 | Gemma 2 27B (IT) | 6.85 | #645 |
| 15 | aya-expanse-8B | 6.85 | #931 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mt-bench-pl-reasoning · How It Works · Data refreshed daily, snapshot 2026-10-20.