UI-Venus-2: General-Purpose GUI Agent for Digital Automation

Multimodal GUI agents hold significant promise for automating digital tasks, but their transition from benchmark-specific models to reliable real-world applications has been hampered by limited environment coverage, brittle task construction, and unreliable reward verification. UI-Venus-2 aims to bridge this gap by presenting itself as a general-purpose foundation GUI agent. This agent is engineered to operate seamlessly across diverse environments, including mobile apps, web interfaces, and native desktop operating systems, all managed through a unified closed-loop reasoning-action framework. The developers focused on scaling three critical dimensions: expanding environment coverage to over 170 multilingual mobile apps and desktop OS; employing a deep-research pipeline for function-grounded instruction generation to create robust tasks; and adopting trace-level and sample-level evaluators with visual keypoints and multi-model voting for reliable reinforcement learning signals. Furthermore, UI-Venus-2 incorporates safety-aware mechanisms to ensure controlled execution of consequential actions, addressing a key concern for real-world deployment. By offering a capable, efficient, and open-source foundation, UI-Venus-2 significantly advances the field towards more generalizable, verifiable, and self-reflective agents suitable for practical applications.