| --- |
| license: mit |
| language: |
| - en |
| tags: |
| - arxiv:2602.16855 |
| --- |
| |
| # reflex |
|
|
| ## Introduction |
|
|
| reflex is a native GUI agent model built on Qwen3-VL. It supports multi-platform GUI automation across desktops, mobile devices, browsers, and more. |
|
|
| * **Paper**: https://arxiv.org/abs/2602.16855 |
| * **GitHub Repository**: https://github.com/X-PLUG/MobileAgent |
| * **Online Demo**: http://modelscope.cn/studios/MobileAgentTest/computer_use |
| |
| **Key highlights:** |
| - 🏆 **State-of-the-art** among multi-platform GUI models on OSWorld-Verified, AndroidWorld, Mobile-World, WindowsAA, ScreenSpot-v2, ScreenSpot-Pro, and more. |
| - 🔧 **Tool & MCP calling**: Native support for external tool invocation and MCP server coordination, achieving top performance on OSWorld-MCP and Mobile-World. |
| - 🧠 **Long-horizon memory**: Built-in memory capability without external workflow orchestration, leading all native agent models on MemGUI-Bench. |
| - 🤝 **Multi-agent ready**: Serves both as a standalone end-to-end agent and as specialized roles (planner, executor, verifier, notetaker) within the Mobile-Agent-v3.5 framework. |
| - ⚡ **Instruct & Thinking variants**: Smaller instruct models for fast inference and edge deployment; larger thinking models for complex tasks requiring planning and reflection. |
| |
| |
| ## Performance |
| ### End-to-End Online Benchmarks |
| |
| | Model | OSWorld-Verified | AndroidWorld | OSWorld-MCP | Mobile-World | WindowsAA | WebArena | VisualWebArena | WebVoyager | Online-Mind2Web |
| |-------|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| |
| | reflex-2B-Instruct | 43.5 | 67.9 | 33.0 | 31.3 | 25.8 | - | - | - | - | |
| | reflex-4B-Instruct | 48.2 | 69.8 | 31.7 | 32.3 | 29.4 | - | - | - | - | |
| | reflex-8B-Instruct | 52.3 | 69.0 | 41.8 | 41.8 | 31.7 | 45.7 | 39.4 | 69.9 | 41.7 | |
| | reflex-8B-Thinking | 52.9 | **71.6** | 38.8 | 33.3 | 35.1 | 46.7 | 40.8 | 78.1 | **48.6** | |
| | reflex-32B-Instruct | **56.5** | 69.4 | **47.6** | **46.8** | **44.8** | - | - | - | - | |
| | reflex-32B-Thinking | 56.0 | 68.2 | 43.8 | 42.8 | 44.1 | **48.4** | **46.6** | **82.1** | - | |
| |
| ### Grounding Benchmarks |
| |
| Please refer to the technical report for detailed results on ScreenSpot-v2, ScreenSpot-Pro, OSWorld-G, MMBench-GUI, and more. |
| |
| ## Usage |
| |
| Please refer to our cookbook. |
| |
| ## Deploy |
| |
| We recommand deploy reflex through vllm |
| |
| This script has been validated on an A100 with 96 GB of VRAM. |
| ```bash |
| PIXEL_ARGS='{"size": {"longest_edge": 3072000, "shortest_edge": 65536}}' |
| IMAGE_LIMIT_ARGS='image=5' |
| MP_SIZE=1 |
| |
| vllm serve $CKPT \ |
| --max-model-len 32768 \ |
| --mm-processor-kwargs "$PIXEL_ARGS" \ |
| --limit-mm-per-prompt "$IMAGE_LIMIT_ARGS" \ |
| --tensor-parallel-size $MP_SIZE \ |
| --allowed-local-media-path '/' \ |
| --port 4243 \ |
| ``` |
| |
| ## Citation |
|
|
| If you find this model useful, please cite our paper: |
|
|
| ```bibtex |
| @article{MobileAgentv3.5, |
| title={Mobile-Agent-v3.5: Multi-platform Fundamental GUI Agents}, |
| author={Haiyang Xu, Xi Zhang, Haowei Liu, Junyang Wang, Zhaozai Zhu, Shengjie Zhou, Xuhao Hu, Feiyu Gao, Junjie Cao, Zihua Wang, Zhiyuan Chen, Jitong Liao, Qi Zheng, Jiahui Zeng, Ze Xu, Shuai Bai, Junyang Lin, Jingren Zhou, Ming Yan}, |
| journal={arXiv preprint arXiv:2602.16855}, |
| year={2026} |
| } |
| |
