MathVision (MATH-V): leaderboard
3,040 math problems with visual contexts from real competitions across 16 disciplines and 5 difficulty levels. Human baseline: 68.8%. Tests multimodal mathematical reasoning from algebra to topology.
Metric: Overall Accuracy (%). Source: mathllm.github.io. 159 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 | 92 | #79 |
| 2 | Gemini 3 Flash | 90.6 | #99 |
| 3 | Gemini 3.1 Pro (Preview) (Thinking) | 89.8 | #58 (Gemini 3.1 Pro (Preview)) |
| 4 | Qwen 3.5 397B A17B | 88.6 | #146 |
| 5 | Kimi K2.6 | 87.4 | #104 |
| 6 | Gemini 3 Pro | 86.6 | #80 |
| 7 | Qwen 3.5 122B A10B | 86.2 | #173 |
| 8 | Qwen 3.5 27B | 86 | #187 |
| 9 | Gemma 4 31B | 85.6 | #200 |
| 10 | Kimi K2.5 | 84.2 | #144 |
| 11 | Qwen 3.5 35B A3B | 83.9 | #256 |
| 12 | GPT-5.2 | 83 | #110 |
| 13 | Gemma 4 26B A4B | 82.4 | #358 |
| 14 | Seed 1.8 | 81.3 | #139 |
| 15 | Seed-1.6 (Thinking) | 77.2 | #265 (Seed-1.6) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mathvision-math-v · How It Works · Data refreshed daily, snapshot 2026-10-20.