X-PLUG/MobileAgent
AI EngineeringA family of GUI agents and GUI-Owl models that automate desktop, mobile, and web apps via screenshots and instructions.
What this is
Mobile-Agent is Tongyi Lab's umbrella repository for GUI automation research and tooling. It spans two connected layers: the GUI-Owl family of multimodal GUI vision-language models, and the Mobile-Agent agent frameworks built on top of them. The agents observe the screen and carry out instructions on desktop, mobile, and browser interfaces, covering planning, progress tracking, reflection, and memory (README).
The current generation is Mobile-Agent-v3.5 / GUI-Owl-1.5, described as native multi-platform agents built on Qwen3-VL, with 2B/4B/8B/32B/235B sizes in Instruct and Thinking variants and claimed state-of-the-art results on more than 20 GUI benchmarks (announcement, technical report). The code for the agent framework lives in the versioned subdirectories rather than at the root (Mobile-Agent-v3.5 code, Mobile-Agent-v3 code).
Older and sibling work is kept in-tree: UI-S1 (semi-online RL for GUI automation, with dataset and model), GUI-Critic-R1 (pre-action error diagnosis), PC-Agent, Mobile-Agent-E, Mobile-Agent-v2, and the original Mobile-Agent. Each has its own paper and subdirectory (series list).
Quick start
The shortest path is the hosted demo, which needs no local model deployment or device setup: open the ModelScope Computer Use demo or the Bailian Computer Use demo, then type a task instruction such as searching for a price or editing a document, as shown in the demo videos for PC/web (example task) and phone (example task).
For programmatic access, Mobile-Agent-v3.5 is available as an API on Alibaba Cloud Bailian; see the API documentation. If you want to run the models yourself, the GUI-Owl-1.5 weights are published on Hugging Face and ModelScope, with 32B and 7B checkpoints linked separately (GUI-Owl-32B, GUI-Owl-7B); the 8B Thinking model is also offered through ModelScope inference (GUI-Owl-1.5-8B-Think).
To run the agent on your own Android device, follow the deployment notes in the Mobile-Agent-v3 directory. For a cloud Android environment instead, the project points to Alibaba Cloud Wuying Cloud Phone and its quick-start guide.
Constraints
- Model and compute: the larger GUI-Owl variants are substantial VLMs; local operation requires the hardware to host or serve them, which is why hosted demos and the Bailian API are the practical entry points.
- Device access: automating a phone requires attaching the agent to a real or cloud Android device (for example via ADB); this is not a pure software-only setup.
- Platform scope: capability claims cover desktop, mobile, and browser, but the code is organized per version and per benchmark (OSWorld, AndroidWorld, real-device scripts); see the OSWorld and AndroidWorld evaluation instructions (OSWorld, AndroidWorld).
- License: the repository is MIT-licensed, but model weights are distributed separately on Hugging Face and ModelScope and may carry their own terms.
- Commercial services: several paths (Bailian API, Wuying Cloud Phone) are Alibaba Cloud products with their own quotas and pricing.
HotRepo's take
There is no HotRepo ranking or star-growth record to report for this repository, so the useful signal is the technical footprint itself: a large, continuously updated research family with published papers, released weights, benchmark code, and hosted demos. The value for readers is less a single tool than a map of the GUI-agent line of work, so it pays to pick one version (currently v3.5) and follow its README rather than the root page.
Written by AI from the project's README, updated . The project's own docs are the source of truth.
You might also like
- kuafuai/andeeAn on-device agent for Android and Mac that operates any app through accessibility, operating apps without integrations or root.
- callstack/agent-deviceA CLI, MCP server, and typed Node.js API that let AI coding agents inspect, control, and verify apps on mobile, TV, web, and desktop devices.
- tapilot-ai/tapilotAn experimental Android app that executes plain-language phone tasks using a vision model and accessibility control.
- minitap-ai/mobile-useAn LLM-driven agent that operates Android and iOS apps via natural language commands and returns structured data.
- AAswordman/OperitAn open-source Android AI agent app that connects cloud or local models to tools, terminals, browsers, and device automation.