Researchers have released Mobile-Agent-v3, a general-purpose GUI agent framework built on the foundational GUI-Owl model, which achieves state-of-the-art performance among open-source end-to-end models across ten GUI benchmarks. The system combines GUI-Owl-7B with a self-evolving trajectory production framework and scalable reinforcement learning to improve automation capabilities.

  • GUI-Owl-7B achieves 66.4 on AndroidWorld and 29.4 on OSWorld, while Mobile-Agent-v3 improves these scores to 73.3 and 37.7 respectively.
  • The framework utilizes a cloud-based virtual environment spanning Android, Ubuntu, macOS, and Windows to generate high-quality interaction data via automated query generation and correctness validation.
  • GUI-Owl integrates UI grounding, planning, action semantics, and reasoning patterns to support end-to-end decision-making as a modular component in multi-agent systems.
  • The team developed Trajectory-aware Relative Policy Optimization (TRPO) for online reinforcement learning, achieving 34.9 on OSWorld with fully asynchronous training.

The open-sourced models provide a self-improving loop that reduces manual annotation requirements and enables real-world alignment for GUI automation tasks.