EvoCUA: Evolving Computer Use Agent

🥇 #1 Open-Source Model on OSWorld | A General-Purpose Multimodal Model Excelling at Computer Use

English | 中文

🥇 #1 Open-Source Model on OSWorld Leaderboard (Jan 2026)

📢 Updates

2026.03.31: EvoCUA-32B achieves 56.48% on WindowsAgentArena (WAA), surpassing the base model Qwen3-VL-32B-Thinking (42.9%) by ~13.6 points and UI-TARS-2 (50.6%) by ~6 points — demonstrating strong zero-shot cross-OS generalization 🆕
2026.03.31: Independent safety study by Yoshua Bengio & Dawn Song's teams (arXiv:2602.08235) shows EvoCUA-32B has the lowest unintended-behavior rate (35.0%) among all tested CUAs — the safest agent! 🆕
2026.01.23: EvoCUA ranked #1 on Hugging Face Daily Papers 🏆
2026.01.22: Released EvoCUA Technical Report 📄
2026.01.13: Released EvoCUA-8B-20260105 — achieves 46.1% on OSWorld, competitive with 72B-level models using fewer parameters!
2026.01.05: Released EvoCUA-32B-20260105 with 56.7% on OSWorld, achieving #1 among open-source models 🥇

🌟 Highlights

🥇 #1 Open-Source Model on OSWorld: Achieves 56.7% task completion rate, #1 among all open-source models
📈 Significant Improvements: +11.7% over OpenCUA-72B (45.0%→56.7%), +15.1% over Qwen3-VL thinking (41.6%→56.7%), with fewer parameters and half the steps
🖥️ End-to-End Multi-Turn Automation: Operates Chrome, Excel, PowerPoint, VSCode and more through screenshots and natural language instructions
🧠 Novel Training Method: Our data synthesis and training approach consistently improves Computer Use capability across multiple open-source VLMs without degrading general performance

📊 Performance Comparison

Rank	Model	Open/Closed	Type	Max Steps	Score
1	Claude-sonnet-4-5	🔒 Closed	General	100	62.9%
2	Seed-1.8	🔒 Closed	General	100	61.9%
3	Claude-sonnet-4-5	🔒 Closed	General	50	58.1%
4	EvoCUA-20260105 (Ours)	🟢 Open	General	50	56.7% 🥇
5	DeepMiner-Mano-72B	🔒 Closed	Specialized	100	53.9%
6	UI-TARS-2-2509	🔒 Closed	General	100	53.1%
7	EvoCUA (Previous Version)	🔒 Closed	General	50	50.3%
8	EvoCUA-8B-20260105 (Ours)	🟢 Open	General	50	46.1%
9	OpenCUA-72B	🟢 Open	Specialized	100	45.0%
...	...	...	...	...	...
13	Qwen3-VL-Flash	🔒 Closed	General	100	41.6%

EvoCUA is #1 among all open-source models, achieving competitive results with only 50 steps. Human-level performance remains significantly higher, indicating substantial room for improvement.

Zero-shot Cross-OS Control (WindowsAgentArena)

We evaluated EvoCUA on WindowsAgentArena (WAA) to test generalization from the Linux-based training environment to a wholly different OS platform. As shown below, EvoCUA-32B reaches 56.48%, surpassing the base model Qwen3-VL-32B-Thinking (42.9%) by ~13.6 points and the leading frontier GUI agent UI-TARS-2 (50.6%) by nearly 6 points.

Model	WAA
Qwen3-VL-32B-Instruct	30.9% [1]
Qwen3-VL-32B-Thinking (Base)	42.9% [1]
UI-TARS-2	50.6% [2]
EvoCUA-32B (Ours)	56.48%

[1] Bai et al., Qwen3-VL Technical Report (arXiv:2511.21631, 2025). [2] Wang et al., UI-TARS-2 Technical Report (arXiv:2509.02544, 2025).

Safety — Robustness to Unintended Behaviors

A recent study by Jones et al. (2026) from Yoshua Bengio and Dawn Song's teams systematically evaluated the safety of leading CUAs. They transferred 117 human-verified perturbations (benign but tricky instructions) to multiple target agents, executing each instruction 3 times and reporting the percentage that elicits unintended behavior in ≥ 1 run. EvoCUA-32B achieves the lowest overall rate (35.0%), demonstrating the strongest robustness among all tested CUAs.

Transferability Study Results (Table 2 from the paper)

Target Agent	Claude 4.5 Haiku	Claude 4.5 Opus	Overall
Open-Source CUAs
EvoCUA-8B	20.0	50.7	37.6
EvoCUA-32B (Ours)	24.0	43.3	35.0 🥇
OpenCUA-7B	42.0	50.7	47.0
OpenCUA-32B	42.0	44.8	43.6
OpenCUA-72B	50.0	56.7	53.8
Closed-Source CUAs
Claude 4.5 Sonnet	32.0	47.8	41.0
Operator	38.0	56.7	48.7

Lower is safer. EvoCUA-32B has the lowest overall unintended-behavior rate among all tested agents.

[3] Jones et al., When Benign Inputs Lead to Severe Harms: Eliciting Unsafe Unintended Behaviors of Computer-Use Agents (arXiv:2602.08235, 2026).

🚀 Quick Start

Installation

Python 3.12 is recommended.

git clone https://github.com/meituan/EvoCUA.git
cd EvoCUA
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Model Download & Deployment

EvoCUA requires downloading the model weights from HuggingFace and deploying with vLLM as an OpenAI-compatible inference server.

Recommended versions:

torch: 2.8.0+cu126
transformers: 4.57.3
vllm: 0.11.0

# 1) Download model weights
huggingface-cli download meituan/EvoCUA-32B-20260105 \
  --local-dir /path/to/EvoCUA-32B \
  --local-dir-use-symlinks False

# 2) Launch vLLM serving (recommend separate environment)
vllm serve /path/to/EvoCUA-32B \
  --served-model-name EvoCUA \
  --host 0.0.0.0 \
  --port 8080 \
  --tensor-parallel-size 2

# 3) Set environment variables
# Environment variables can be configured in .env file (see env.template for reference):
cp env.template .env
# Edit .env with your configurations, e.g.,
export OPENAI_API_KEY="dummy"
export OPENAI_BASE_URL="http://127.0.0.1:8080/v1"

Run Evaluation on OSWorld

python3 run_multienv_evocua.py \
  --headless \
  --provider_name aws \
  --observation_type screenshot \
  --model EvoCUA-S2 \
  --result_dir ./evocua_results \
  --test_all_meta_path evaluation_examples/test_nogdrive.json \
  --max_steps 50 \
  --num_envs 30 \
  --temperature 0.01 \
  --max_history_turns 4 \
  --coordinate_type relative \
  --resize_factor 32 \
  --prompt_style S2

📁 Project Structure

EvoCUA/
├── run_multienv_evocua.py      # Main entry point (multi-env parallel evaluation)
├── lib_run_single.py           # Single task rollout logic (trajectory, screenshots, recording, scoring)
├── lib_results_logger.py       # Real-time result aggregation to results.json
├── desktop_env/                # OSWorld environment implementation
│   ├── providers/              # VM providers (AWS/VMware/Docker/etc.)
│   ├── controllers/            # Environment controllers
│   └── evaluators/             # Task evaluators
├── mm_agents/
│   └── evocua/                 # EvoCUA agent (prompts, parsing, action generation)
└── evaluation_examples/        # OSWorld task configurations

📖 About OSWorld

OSWorld is the most influential benchmark in the Computer Use Agent domain. It is adopted by leading AI organizations including OpenAI, Anthropic, ByteDance Seed, Moonshot AI, Zhipu AI, Step, and more. OSWorld evaluates agents' ability to complete real-world computer tasks through multi-turn interactions with actual desktop environments.

🔗 Resources

🤗 Model Weights:
- meituan/EvoCUA-32B-20260105 - OSWorld Score: 56.7% 🥇
- meituan/EvoCUA-8B-20260105 - OSWorld Score: 46.06% 🆕
📊 OSWorld Benchmark: os-world.github.io
📄 Technical Report: tech_report.pdf
🚀 More Model Sizes: More models of various sizes are on the way!

🙏 Acknowledgements

We sincerely thank the open-source community for their outstanding contributions to the Computer Use Agent field. We are grateful to Xinyuan Wang (OpenCUA) and Tianbao Xie (OSWorld) for their insightful discussions, valuable feedback on evaluation, and continuous support throughout this project. Their pioneering work has greatly inspired and advanced our research. We are committed to giving back to the community and will continue to open-source our research to advance the field.

📝 Citation

If you find EvoCUA useful in your research, please consider citing:

@article{xue2026evocua,
  title={EvoCUA: Evolving Computer Use Agents via Learning from Scalable Synthetic Experience},
  author={Xue, Taofeng and Peng, Chong and Huang, Mianqiu and Guo, Linsen and Han, Tiancheng and Wang, Haozhe and Wang, Jianing and Zhang, Xiaocheng and Yang, Xin and Zhao, Dengchang and others},
  journal={arXiv preprint arXiv:2601.15876},
  year={2026}
}