We-Math
Multimodal · 2024-07-01
We-Math evaluates visual mathematical reasoning with problems decomposed into sub-problems over 67 hierarchical knowledge concepts, diagnosing insufficient knowledge, inadequate generalization and rote memorization. This row holds the single headline score reported in model releases, which do not always state whether it is the strict or loose metric.
Top models (higher is better)
| Model | Score |
|---|---|
| Qwen3.6 Plus (2026-04-02) | 89.0 |
| Qwen3.5 397B A17B | 87.9 |
| Gemini 3 Pro Preview | 86.9 |
| Kimi K2.5 | 84.7 |
| GPT-5.2 | 79.0 |
| Qwen3.5-4B | 75.4 |
| Qwen3.5-9B | 75.2 |
| Opus 4.5 | 70.0 |
| GPT-5 Nano | 62.5 |
| Gemini 2.5 Flash-Lite | 32.1 |