Qwen2.5 VL 32B Instruct
Qwen2.5-VL is a vision-language model from the Qwen family. Key enhancements include visual understanding (objects, text, charts, layouts), visual agent capabilities (tool use, computer/phone control), long video comprehension with event pinpointing, visual localization (bounding boxes/points), and structured output generation.
Benchmark results
| Benchmark | Score | Tags | Source |
|---|---|---|---|
| AITZ_EM | 83.1% | self-reported llm-stats | link → |
| Android Control High_EM | 69.6% | self-reported llm-stats | link → |
| Android Control Low_EM | 93.3% | self-reported llm-stats | link → |
| AndroidWorld_SR | 22.0% | self-reported llm-stats | link → |
| CC-OCR | 77.1% | self-reported llm-stats | link → |
| CharadesSTA | 54.2% | self-reported llm-stats | link → |
| DocVQA | 94.8% | self-reported llm-stats | link → |
| GPQA | 46.0% | self-reported llm-stats | link → |
| HumanEval | 91.5% | self-reported llm-stats | link → |
| InfoVQA | 83.4% | self-reported llm-stats | link → |
| LVBench | 49.0% | self-reported llm-stats | link → |
| MATH | 82.2% | self-reported llm-stats | link → |
| MathVision | 38.4% | self-reported llm-stats | link → |
| MathVista-Mini | 74.7% | self-reported llm-stats | link → |
| MBPP | 84.0% | self-reported llm-stats | link → |
| MMBench-Video | 1.9% | self-reported llm-stats | link → |
| MMLU | 78.4% | self-reported llm-stats | link → |
| MMLU-Pro | 68.8% | self-reported llm-stats | link → |
| MMMU | 70.0% | self-reported llm-stats | link → |
| MMMU-Pro | 49.5% | self-reported llm-stats | link → |
| MMStar | 69.5% | self-reported llm-stats | link → |
| OCRBench-V2 (en) | 57.2% | self-reported llm-stats | link → |
| OCRBench-V2 (zh) | 59.1% | self-reported llm-stats | link → |
| OSWorld | 5.9% | self-reported llm-stats | link → |
| ScreenSpot | 88.5% | self-reported llm-stats | link → |
| ScreenSpot Pro | 39.4% | self-reported llm-stats | link → |
| VideoMME w sub. | 77.9% | self-reported llm-stats | link → |
| VideoMME w/o sub. | 70.5% | self-reported llm-stats | link → |