DynaMath
Multimodal · 2024-10-29
DynaMath is a dynamic visual math benchmark that generates 10 variants of each of 501 seed questions by programmatically changing values, functions or graphs, for 5,010 questions. Release tables report average accuracy across all generated variants; worst-case accuracy is a separate metric not stored here.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.6 Plus (2026-04-02) | 88.0 |
| Qwen3.5-27B | 87.7 |
| Gemma 4 12B IT | 87.6 |
| GPT-5.2 | 86.8 |
| Qwen3.5 397B A17B | 86.3 |
| Qwen3.5 122B A10B | 85.9 |
| Qwen3.6 27B | 85.6 |
| Gemini 3 Pro Preview | 85.1 |
| Qwen3.5 35B A3B | 85.0 |
| GPT-5.6 Luna | 84.6 |
| Qwen3.5-4B | 84.6 |
| Kimi K2.5 | 84.4 |
| Qwen3.5 Plus (2026-02-15) | 84.2 |
| Qwen3.5-Omni-Plus | 83.8 |
| Qwen3.5-9B | 83.6 |
| Qwen3.6 35B A3B | 82.8 |
| Gemini 3.5 Flash-Lite | 82.2 |
| GPT-5 Mini | 81.4 |
| Opus 4.5 | 79.7 |
| Gemma 4 31B IT | 79.5 |
| Qwen3.5-Omni-Flash | 79.3 |
| LFM3-VL-2B | 79.2 |
| Sonnet 4.5 | 78.8 |
| GPT-5 Nano | 78.0 |
| Qwen3.5-2B | 73.6 |
| Gemini 2.5 Flash-Lite | 69.9 |
| Gemma 4 E2B IT | 63.9 |
| Qwen3.5-0.8B | 49.9 |