Math-VR: leaderboard
Mathematical visual reasoning benchmark for VLMs, unified models, and LLMs, reporting answer correctness and process scores on text and multimodal questions.
Metric: Overall Answer Correctness (imported). Source: huggingface.co. Source: benchmarklist.com. Saturation forecast: Around February 2027. 30 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 VL 235B A22B (Thinking) | 66.8 | #238 (Qwen 3 VL 235B A22B) |
| 2 | Qwen 3 VL 235B A22B Instruct | 65 | #273 |
| 3 | Gemini 2.5 Pro | 64.7 | #152 |
| 4 | Gemini 2.5 Flash | 60.5 | #245 |
| 5 | O3 | 59.3 | #124 |
| 6 | Seed-1.6 (Thinking) | 58.4 | #265 (Seed-1.6) |
| 7 | GPT-5 (Thinking) | 58.1 | #98 (GPT-5) |
| 8 | Claude Opus 4.1 | 54.3 | #140 |
| 9 | Gemini 2.5 Flash (Non-reasoning) | 52.3 | #245 (Gemini 2.5 Flash) |
| 10 | GLM-4.5V | 49.6 | #346 |
| 11 | GPT-4.1 Mini | 33.3 | #355 |
| 12 | Claude Sonnet 4 | 28.1 | #199 |
| 13 | GPT-4.1 | 26 | #248 |
| 14 | Gemini 2.0 Flash | 20.6 | #342 |
| 15 | GPT-4.1 Nano | 9.1 | #731 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=math-vr · How It Works · Data refreshed daily, snapshot 2026-10-20.