Add model card and metadata for ClawGUI-2B

README.md

@@ -1,3 +1,31 @@
 ---
 license: apache-2.0
+pipeline_tag: image-text-to-text
+library_name: transformers
 ---
+
+# ClawGUI-2B
+
+**ClawGUI-2B** is a 2B-parameter multimodal GUI agent trained as part of the [ClawGUI framework](https://github.com/zju-real/ClawGUI). It is designed to interact with software through visual interfaces by predicting actions such as taps, swipes, and keystrokes directly from screenshots and natural language instructions.
+
+## Resources
+- **Paper:** [ClawGUI: A Unified Framework for Training, Evaluating, and Deploying GUI Agents](https://huggingface.co/papers/2604.11784)
+- **GitHub Repository:** [ZJU-REAL/ClawGUI](https://github.com/zju-real/ClawGUI)
+- **Project Page:** [https://zju-real.github.io/ClawGUI-Page](https://zju-real.github.io/ClawGUI-Page)
+
+## Model Description
+ClawGUI-2B is trained end-to-end within the ClawGUI pipeline using **GiGPO** (a variation of GRPO) and a **Process Reward Model (PRM)** to provide dense step-level supervision. This training methodology allows the agent to reach high levels of stability and performance across different environments, including virtual emulators and real physical devices.
+
+In benchmark evaluations, ClawGUI-2B achieves a **17.1% Success Rate** on the MobileWorld GUI-Only benchmark, outperforming the MAI-UI-2B baseline of the same scale by 6.0%.
+
+## Citation
+If you find ClawGUI useful in your research, please cite:
+
+```bibtex
+@article{tang2026clawgui,
+  title={ClawGUI: A Unified Framework for Training, Evaluating, and Deploying GUI Agents},
+  author={Fei Tang and Zhiqiong Lu and Boxuan Zhang and Weiming Lu and Jun Xiao and Yueting Zhuang and Yongliang Shen},
+  journal={arXiv preprint arXiv:2604.11784},
+  year={2026}
+}
+```