Downloads · 30 days
178
1% of all-time downloads
meituan/EvoCUA-32B-20260105
EvoCUA-32B-20260105 is a machine learning model from meituan. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
Downloads · 30 days
178
1% of all-time downloads
All-time downloads
12.9K
Public
Parameters
33.4B
66.8 GB on disk
Likes
33
Trending 1
Click a slice to open those files.
.safetensors66.7 GB · 100%
From the Hugging Face model README
🥇 #1 Open-Source Model on OSWorld | A General-Purpose Multimodal Model Excelling at Computer Use
<img src="assets/images/osworld_leaderboard.png" width="700" alt="OSWorld Leaderboard">🥇 #1 Open-Source Model on OSWorld Leaderboard (Jan 2026)
</div>| 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.
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).
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.
<details> <summary><b>Transferability Study Results (Table 2 from the paper)</b></summary>| 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 |
</details>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).
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
EvoCUA requires downloading the model weights from HuggingFace and deploying with vLLM as an OpenAI-compatible inference server.
Recommended versions:
# 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"
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
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
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.
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.
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}
}
This project is licensed under the Apache 2.0 License - see the LICENSE file for details.
Built with ❤️ by Meituan LongCat Team
</div>