MAI-UI: Alibaba’s New Foundation GUI Agents Outperforming Gemini & GPT-4o

The next generation of human-computer interaction has arrived with MAI-UI, a family of foundation GUI agents designed to perceive, reason, and act within digital interfaces. Developed by Alibaba’s Tongyi Lab, MAI-UI transforms manual navigation into goal-oriented natural language control. In this video, we dive into the technical breakthroughs of the MAI-UI family, which spans from 2B on-device models to massive 235B-A22B variants. Unlike previous agents, MAI-UI addresses the critical gaps required for real-world deployment, including native agent-user interaction, MCP tool integration, and a pioneering device-cloud collaboration system. Key Highlights of MAI-UI: • New State-of-the-Art Performance: MAI-UI establishes new records across five grounding benchmarks and mobile navigation, achieving a 76.7% success rate on AndroidWorld, surpassing UI-Tars-2, Gemini-2.5-Pro, and Seed1.8. • Device-Cloud Collaboration: A local agent acts as both an executor and a monitor, routing complex tasks to a high-capacity cloud agent while preserving user privacy by keeping sensitive data (like passwords) on-device. • Beyond UI-Only Actions: By integrating the Model Context Protocol (MCP), MAI-UI can compress long UI sequences into efficient API calls, enabling desktop-level workflows like GitHub repository manipulation on mobile devices. • Robustness via Online RL: Using an advanced GRPO reinforcement learning framework, the agent is trained in over 512 parallel dynamic environments, making it resilient to unexpected pop-ups and permission dialogs. • Instruction-as-Reasoning: The model is trained to think through different perspectives (appearance, function, location, and intent) before acting, which prevents "policy collapse" and enhances grounding accuracy. Whether you're interested in the future of mobile automation or the latest in Multimodal Large Language Models (MLLMs), MAI-UI represents a significant step toward practical, reliable AI executors https://tongyi-mai.github.io/MA