UI-Venus-1.5 技术报告

UI-Venus-1.5 is a unified, end-to-end GUI agent designed for robust real-world applications. The model family includes two dense variants (2B and 8B) and one mixture-of-experts variant (30B-A3B) to cater to various downstream application scenarios. Compared to its predecessor, UI-Venus-1.5 introduces three key technical advancements: first, a comprehensive Mid-Training stage leveraging 10 billion tokens across 30+ datasets to establish foundational GUI semantics; second, the implementation of Online Reinforcement Learning with full-trajectory rollouts to enhance the model's learning capabilities; and third, optimizations in inference speed and accuracy, improving performance on complex tasks. This research lays a significant technical foundation for future automation in graphical user interfaces.