PolyMath
Math · 2025-04-25
PolyMath is a multilingual mathematical reasoning benchmark of 9,000 problems: 125 problems at each of four difficulty levels, from grade school to Olympiad, translated into 18 languages. Its official headline metric, difficulty-weighted accuracy, weights the four levels 1, 2, 4 and 8 before averaging across languages; release tables that label it simply PolyMath report this same figure. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.7-Max (source unspecified) | 86.5 |
| Qwen3.7-Plus (source unspecified) | 84.0 |
| Kimi K2.6 | 82.7 |
| Gemini 3 Pro Preview | 81.6 |
| Opus 4.6 | 80.2 |
| Opus 4.5 | 79.0 |
| Qwen3.6 Plus (2026-04-02) | 77.4 |
| Qwen3.5 397B A17B | 73.3 |
| DeepSeek-V4-Pro | 72.0 |
| Qwen3.5-27B | 71.2 |
| Qwen3.5 122B A10B | 68.9 |
| GLM-5.1 | 67.6 |
| GPT-5 Mini | 67.3 |
| GLM-5 | 65.2 |
| Qwen3 Max Thinking (2026-01-23) | 64.7 |
| Qwen3.5 35B A3B | 64.4 |
| GPT-5.2 | 62.5 |
| Qwen3-235B-A22B | 60.1 |
| Qwen3 235B A22B Thinking 2507 | 60.1 |
| Qwen3 VL 235B A22B Thinking | 57.8 |
| Qwen3.5-9B | 57.3 |
| gpt-oss-120b | 54.0 |
| Qwen3 30B A3B Thinking 2507 | 52.6 |
| Gemini 2.5 Pro | 52.2 |
| Qwen3 VL 32B Thinking | 52.0 |
| Qwen3 VL 30B A3B Thinking | 51.7 |
| Qwen3.5-4B | 51.1 |
| Qwen3 235B A22B Instruct 2507 | 50.2 |
| Gemini 2.5 Flash | 49.8 |
| o3 | 49.7 |
| O4 Mini | 48.7 |
| Qwen3 VL 8B Thinking | 47.5 |
| Qwen3 32B | 47.4 |
| Qwen3 Omni 30B A3B Thinking | 47.1 |
| DeepSeek-R1 | 47.1 |
| DeepSeek-R1-0528 | 46.9 |
| Qwen3 4B Thinking 2507 | 46.2 |
| Qwen3-30B-A3B | 46.1 |
| Qwen3 Next 80B A3B Instruct | 45.9 |
| QwQ-32B | 45.9 |