metadata
license: apache-2.0
tags:
- Reinforcement Learning
- Visual-langauge Reasoning
Model Card for WeThink-Qwen2.5VL-7B
Repository: https://github.com/yangjie-cv/WeThink
Paper: https://arxiv.org/abs/2506.07905
🏆 Performance Highlights
WeThink-Qwen2.5VL-7B achieves:
- 🥇 1st place on OpenCompass Multimodal Reasoning Leaderboard
- 🏅 5th place on OpenCompass Multi-modal Academic Leaderboard
(As of May 30th, 2025)
🚀 Quick Start
Inference
git clone https://github.com/yangjie-cv/WeThink
cd WeThink
python inference.py
💡 Note: System prompt is required during inference.
📊 Evaluation
We have integrated WeThink-Qwen2.5VL-7B into the VLMEvalKit. Please follow its Quickstart guide to evaluate WeThink-Qwen2.5VL-7B on various benchmarks.
Citation
@misc{yang2025wethink,
title={WeThink: Toward General-purpose Vision-Language Reasoning via Reinforcement Learning},
author={Jie Yang and Feipeng Ma and Zitian Wang and Dacheng Yin and Kang Rong and Fengyun Rao and Ruimao Zhang},
year={2025},
eprint={2506.07905},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2506.07905},
}