MobileAgent (Mobile-Agent) is a series of GUI Agent projects launched by Alibaba Group's Tongyi Lab, aiming to build a family of AI agents capable of autonomously operating graphical user interfaces. The project is hosted on GitHub and continuously releases multiple versions and accompanying models.
[Purpose] The Mobile-Agent series is used to enable AI agents to perform multi-step GUI operation tasks on platforms such as phones, computers, and browsers. For example: searching for stock prices and creating a spreadsheet in WPS Office to fill in the data; creating a document in WPS Office and inserting an image copied from a browser; searching for flights and train tickets on Ctrip and comparing prices; checking the total number of followers of Xiaohongshu and Douyin accounts, etc. These examples demonstrate its cross-application and cross-platform automation capabilities.
[Main Capabilities]
- Multi-platform support: covers desktop (PC), mobile (Android), and browser automation.
- Multi-agent framework: Mobile-Agent-v3 is a cross-platform multi-agent framework based on GUI-Owl, providing capabilities such as planning, progress management, reflection, and memory.
- GUI-Owl foundation models: GUI-Owl is a multimodal cross-platform GUI VLM with GUI perception, grounding, and end-to-end operation capabilities. The GUI-Owl-1.5 series is built on Qwen3-VL, offering parameter scales of 2B/4B/8B/32B/235B, including Instruct and Thinking versions, supporting desktop/mobile/browser automation, achieving SOTA results on more than 20 GUI benchmarks, and performing strongly in end-to-end tasks, grounding, tool/MCP calling, and long-term memory.
- Tool calling and MCP: supports tool/MCP calling, and has released the OSWorld-MCP benchmark for evaluating MCP tool calling capabilities in real-world scenarios.
- ToolCUA: an end-to-end Computer Use Agent for optimizing GUI-tool path orchestration, adopting a two-stage training process (trajectory-aware tool synthesis -> online agent reinforcement learning), mastering when to use GUI operations, call tools, and switch.
- Other research results: UI-S1 (semi-online reinforcement learning advancing GUI automation), GUI-Critic-R1 (a GUI critic for pre-operative error diagnosis), PC-Agent (a multimodal PC operation multi-agent), Mobile-Agent-E (a self-evolving mobile operation multi-agent), etc.
[Usage]
- Online experience: Mobile-Agent-v3.5 can be tried through the ModelScope online demo or the Alibaba Cloud Bailian online demo.
- API: Alibaba Cloud Bailian provides the Mobile-Agent-v3.5 API, and you can check the documentation for a quick experience; previously, a limited-time free Mobile-Agent-v3 API was also provided.
- Cloud phone: Mobile-Agent-v3.5 has been launched on Alibaba Cloud Wuying Cloud Phone, providing a cloud-based Android environment.
- Model weights: The GUI-Owl-1.5 series model weights can be obtained on HuggingFace, and the technical report can be found via the link.
- Open-source code: The code for GUI-Owl and Mobile-Agent-v3 on OSWorld, AndroidWorld, and real mobile scenarios has been open-sourced, including the RL tuning checkpoint for OSWorld.
- Online inference: The GUI-Owl-1.5 series models can perform online inference through Alibaba Cloud Bailian and ModelScope API-Inference.
[Applicable Scenarios]
- Mobile automation: executing cross-application tasks on Android devices.
- Desktop automation: completing operations such as document editing and data entry on a PC.
- Browser automation: web search, information extraction, image copying, etc.
- Research and development: academic research related to GUI agents, benchmarking, model training, and evaluation.
[Limitations] Regarding specific fees, prices, platform compatibility details, performance metrics, etc., the materials do not provide complete information and are not yet available. Some features may depend on specific platforms such as Alibaba Cloud Bailian, ModelScope, or Wuying Cloud Phone.
