Hugging Face has released a technical report introducing UI-Venus-2, a general-purpose multimodal GUI agent designed to handle digital tasks across mobile, web, and desktop platforms. The agent employs a unified closed-loop reasoning-action framework, aiming to bridge the gap between benchmark-oriented models and dependable real-world applications.
The development of UI-Venus-2 focuses on scaling three critical dimensions simultaneously. First, it expands environment coverage to more than 170 multilingual mobile apps and native desktop operating systems. Second, it uses a deep-research pipeline to generate function-grounded instructions for task construction. Third, it adopts trace-level and sample-level evaluators, incorporating visual keypoints and multi-model voting to ensure reliable reinforcement learning signals during training.
To ensure safe deployment, UI-Venus-2 integrates safety-aware mechanisms that enable controlled execution of consequential actions. The agent is positioned as an open-source foundation that is both capable and efficient, advancing the field toward more generalizable, verifiable, and self-reflective agents for real-world applications.