Benchmark
MathVision Benchmark
What is MathVision?
MATH-Vision is a dataset designed to measure multimodal mathematical reasoning capabilities. It focuses on evaluating how well models can solve mathematical…
- Released
- 2024-02-22
- Modality
- multimodal
- Evaluates
- math, multimodal, vision
- Openness
- open
- Importer
- LLM Stats
- Reported scores
- 33
Source provenance
- Original evidence https://api.zeroeval.com/leaderboard/benchmarks/mathvision?top_n=500
MathVision paper, code and dataset
MathVision results and reported scores
LLM Stats
| Model | Organization | Reported value | Reported | Evidence |
|---|---|---|---|---|
| DiffusionGemma 26B-A4B | 0.705 | 2026-06-10 | evidence | |
| Gemma 4 12B | 0.797 | 2026-05-23 | evidence | |
| Gemma 4 26B-A4B | 0.824 | 2026-04-02 | evidence | |
| Gemma 4 31B | 0.856 | 2026-04-02 | evidence | |
| Gemma 4 E2B | 0.524 | 2026-04-02 | evidence | |
| Gemma 4 E4B | 0.595 | 2026-04-02 | evidence | |
| Kimi K2.5 | Moonshot AI | 0.842 | 2026-01-27 | evidence |
| Kimi K2.6 | Moonshot AI | 0.932 | 2026-04-20 | evidence |
| Kimi K3 | Moonshot AI | 0.978 | 2026-07-16 | evidence |
| QvQ-72B-Preview | Qwen | 0.359 | 2024-12-25 | evidence |
| Qwen2.5 VL 32B Instruct | Qwen | 0.384 | 2025-02-28 | evidence |
| Qwen2.5 VL 72B Instruct | Qwen | 0.381 | 2025-01-26 | evidence |
| Qwen2.5 VL 7B Instruct | Qwen | 0.2507 | 2025-01-26 | evidence |
| Qwen2.5-Omni-7B | Qwen | 0.25 | 2025-03-27 | evidence |
| Qwen3 VL 235B A22B Instruct | Qwen | 0.665 | 2025-09-22 | evidence |
| Qwen3 VL 235B A22B Thinking | Qwen | 0.746 | 2025-09-22 | evidence |
| Qwen3 VL 30B A3B Instruct | Qwen | 0.602 | 2025-09-22 | evidence |
| Qwen3 VL 30B A3B Thinking | Qwen | 0.657 | 2025-09-22 | evidence |
| Qwen3 VL 32B Instruct | Qwen | 0.634 | 2025-09-22 | evidence |
| Qwen3 VL 32B Thinking | Qwen | 0.702 | 2025-09-22 | evidence |
| Qwen3 VL 4B Instruct | Qwen | 0.516 | 2025-09-22 | evidence |
| Qwen3 VL 4B Thinking | Qwen | 0.6 | 2025-09-22 | evidence |
| Qwen3 VL 8B Instruct | Qwen | 0.539 | 2025-09-22 | evidence |
| Qwen3 VL 8B Thinking | Qwen | 0.627 | 2025-09-22 | evidence |
| Qwen3.5-122B-A10B | Qwen | 0.862 | 2026-02-24 | evidence |
| Qwen3.5-27B | Qwen | 0.86 | 2026-02-24 | evidence |
| Qwen3.5-35B-A3B | Qwen | 0.839 | 2026-02-24 | evidence |
| Qwen3.6 Plus | Qwen | 0.88 | 2026-04-02 | evidence |
| Qwen3.7-Plus | Qwen | 0.903 | 2026-05-31 | evidence |
| Qwen3.8-27B | Qwen | 0.946 | 2026-08-14 | evidence |
| Seed 2.1 Pro | ByteDance | 0.945 | 2026-06-24 | evidence |
| Seed 2.1 Turbo | ByteDance | 0.927 | 2026-06-24 | evidence |
| Step3-VL-10B | StepFun | 0.708 | 2026-01-15 | evidence |
Scores are partitioned by the source that reported them and are never merged into a single cross-source ranking, because the sources measure different things and say so.