---
language:
- en
pretty_name: WebHarbor-4B
library_name: sglang
pipeline_tag: image-text-to-text
base_model: OpenWebRL/OpenWebRL-4B-SFT
datasets:
- ChilleD/WebHarbor
tags:
- webharbor
- web-agent
- gui-agent
- computer-use-agents
- reinforcement-learning
- qwen3-vl
---

<h1 align="center">WebHarbor: Docking Real Websites for Evolving GUI Agent Environments</h1>

<h2 align="center">WebHarbor-4B Model</h2>

<p align="center">
  <a href="https://aiming-lab.github.io/WebHarbor/"><img src="https://img.shields.io/badge/Project-Page-0A7EA4.svg" alt="Project Page"></a>
  <a href="https://arxiv.org/abs/2611.03634"><img src="https://img.shields.io/badge/arXiv-2611.03634-b31b1b.svg" alt="arXiv"></a>
  <a href="https://github.com/aiming-lab/WebHarbor"><img src="https://img.shields.io/badge/Code-GitHub-181717.svg" alt="Code"></a>
  <a href="https://huggingface.co/datasets/ChilleD/WebHarbor"><img src="https://img.shields.io/badge/🤗-153_Environments-yellow.svg" alt="153 Environments"></a>
  <a href="https://hub.docker.com/r/wildawm/webharbor"><img src="https://img.shields.io/badge/Docker-wildawm%2Fwebharbor-2496ED.svg" alt="Docker Images"></a>
</p>

<p align="left">
  <b>WebHarbor-4B</b> is a GUI web agent trained with reinforcement learning entirely inside <a href="https://aiming-lab.github.io/WebHarbor/">WebHarbor</a>, 153 executable reconstructions of real websites with logins, carts, bookings and 3,518 tasks. It never touches the live web during RL training, yet it transfers there: it reaches a <b>75.8</b> average on WebVoyager and Online-Mind2Web, the best among the open-source web agents compared. The model is a 4B <a href="https://huggingface.co/Qwen/Qwen3-VL-4B-Thinking">Qwen3-VL</a> policy that starts from OpenWebRL-4B-SFT and acts through the <a href="https://github.com/OpenWebRL/OpenWebRL">OpenWebRL</a> browser interface.
</p>

---

## 🔮 Resources

| Resource | Link |
|---|---|
| 🌐 Project page | [WebHarbor](https://aiming-lab.github.io/WebHarbor/) |
| 📄 Paper | [arXiv:2611.03634](https://arxiv.org/abs/2611.03634) |
| 💻 Code (environments, evaluation) | [WebHarbor repository](https://github.com/aiming-lab/WebHarbor) |
| 📦 Environments | [ChilleD/WebHarbor](https://huggingface.co/datasets/ChilleD/WebHarbor) |
| 🐳 Docker images | [wildawm/webharbor](https://hub.docker.com/r/wildawm/webharbor) |
| 🤖 WebHarbor-4B | [ChilleD/WebHarbor-4B](https://huggingface.co/ChilleD/WebHarbor-4B) |

## ✨ Highlights

| Benchmark | OpenWebRL-4B-SFT | OpenWebRL-4B | **WebHarbor-4B** |
|---|:---:|:---:|:---:|
| WebVoyager | 60.2 | 74.1 | **81.2** |
| Online-Mind2Web | 47.0 | 67.0 | **70.4** |
| WebArena | 26.1 | 20.4 | **28.8** |
| WebArena without auto-login | 21.2 | 16.8 | **26.1** |
| WebHarbor test set | 29.1 | 19.7 | **30.0** |

- **Live web from offline training:** a 75.8 average on WebVoyager and Online-Mind2Web, ahead of GPT-5 (74.2) and every listed open-source agent.
- **Strongest on real-world workflows:** the largest gains are on logins, multi-page navigation, stateful operations and transactions, the workflows that live-web training rarely covers.
- **Unseen sites and tasks:** all 15 WebVoyager websites are held out of training, and Online-Mind2Web and WebArena tasks are not in the training data.

## 📦 Setup

```bash
uv pip install "sglang==0.5.6.post2"
```

## 🔥 Serving

A single-GPU launch with the 32K context used in our evaluation:

```bash
python -m sglang.launch_server \
  --model-path ChilleD/WebHarbor-4B \
  --served-model-name webharbor-4b \
  --host 0.0.0.0 --port 30000 \
  --dtype bfloat16 \
  --context-length 32768 \
  --mem-fraction-static 0.9 \
  --trust-remote-code \
  --attention-backend flashinfer \
  --mm-attention-backend triton_attn
```

This exposes an OpenAI-compatible Chat Completions endpoint at `http://localhost:30000/v1`. The chat template already opens the `<think>` block, so clients should not add their own. For the browser harness and to evaluate on WebHarbor, see the [WebHarbor repository](https://github.com/aiming-lab/WebHarbor).

## Citation

If you find this work useful, please cite:

```bibtex
@article{wang2026webharbordockingrealwebsites,
      title={WebHarbor: Docking Real Websites for Evolving GUI Agent Environments}, 
      author={{WebHarbor Team and Contributors}},
      year={2026},
      eprint={2611.03634},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2611.03634}, 
}
```

or

```bibtex
@article{wang2026webharbordockingrealwebsites,
      title={WebHarbor: Docking Real Websites for Evolving GUI Agent Environments}, 
      author={Zhaoyang Wang and Qianhui Wu and Boyu Gou and Demi Wang and Jiacheng Zhang and Peichun Hua and Qize Li and Shi Qiu and Yifan Jiang and Yifan Zhang and Zexu Jin and Baolin Peng and Bolin Shen and Chetan Bansal and Dehan Kong and Deren Lei and Dongshen Peng and Fengqing Jiang and Hao Cheng and Haoqing Wang and Jiaming Shan and Jicheng Wang and Jingguang Zhou and Jixuan He and Kaide Zeng and Kangyu Zhu and Lama Moukheiber and Mufan Qiu and Peng Xia and Qihan Sun and Qingwei Lin and Qiushi Sun and Ruichen Zhang and Shuqing Luo and Si Qin and Sicheng Fan and Tao Xiong and Wenlin Yao and Xinyu Zhang and Xinyue Zeng and Xuanrui Li and Xuchao Zhang and Yian Wang and Yichao Liu and Yiming Liang and Yiyang Zhou and Yizhu Jiao and Yuetai Li and Yurun Chen and Yuxuan Zhang and Zheyuan Yang and Zhiyuan Yao and Zhongyang Li and Zihao Li and Zili Wang and Jianfeng Gao and Huaxiu Yao},
      year={2026},
      eprint={2611.03634},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2611.03634}, 
}
```