Tongyi-MAI/MAI-UI

What it solves

Qwen-UI-Agent (the evolution of MAI-UI) is a foundation GUI agent designed to operate across mobile devices, computers, web browsers, and DeepSearch environments. It solves the problem of creating a single model that can autonomously navigate and execute complex, long-horizon tasks across different platforms and interfaces, closing the gap between simulated environments and real-world device usage.

How it works

The agent utilizes a hybrid action space that combines GUI operations (clicks, etc.) with CLI (Bash) commands, allowing it to perform both visual navigation and direct system-level execution. It is trained using scalable long-horizon online reinforcement learning (RL) across thousands of parallel environments. To improve its capabilities, it employs an "AutoResearch-style" data flywheel where the agent itself helps construct tasks, diagnose failures, and plan iterations. Additionally, a harness layer allows it to be proactive, initiating tasks based on real-world signals like notifications.

Who it’s for

Researchers and developers building autonomous agents capable of cross-platform GUI navigation, system automation, and real-world device interaction.

Highlights

• Unified Interface: Operates across mobile, PC, and web in a single model.
• Real-Device Training: Trained and evaluated on a live environment of 100+ physical smartphones.
• Hybrid Action Space: Combines GUI clicks with Bash commands and supports batched actions per turn.
• Proactive Service: Can trigger workflows based on external signals (e.g., flight cancellation notifications).
• Cross-Platform Workflows: Capable of moving data and executing tasks between mobile and computer devices.
• High Performance: Achieves state-of-the-art results across multiple benchmarks including MobileWorld, AndroidDaily, and OSWorld.