Ant's UI-Venus-1.5 Tops SOTA for GUI Agents

• Architecture Foundation: Built on Qwen3-VL, a large vision-language model, adapted specifically for GUI understanding and navigation tasks • Multi-Stage Training Pipeline: (1) Mid-training phase with 10 billion tokens across 30+ datasets for GUI semantic understanding; (2) Online reinforcement learning with full-trajectory rollouts for long-horizon task alignment • Model Merge Strategy: Synthesizes three specialized capabilities—grounding, web navigation, and mobile interaction—into a single unified checkpoint through intelligent model merging • Platform Agnosticism: Operates at pixel level, decoupling AI from operating system specifics, enabling operation across native applications, web interfaces, and remote desktops • Performance Variants: Available in 2B (8.7% VenusBench-Mobile success), 8B (16.1%), and 30B-A3B (21.5%) parameter scales with proportional performance improvements • Benchmark Coverage: Evaluated on AndroidWorld (77.6%), AndroidLab (55.1%/68.1%), OSWorld-G-R (76.4%), OSWorld-G (70.6%), VenusBench-Mobile (21.5%), and WebVoyager (76.0%) • Key Innovation: Addresses step-level vs. task-level accuracy disparity by optimizing full interaction sequences rather than individual steps, improving real-world task completion rates

UI-Venus-1.5 represents a significant advancement in autonomous GUI agents with implications across multiple sectors. The unified, end-to-end architecture establishes a new performance ceiling for GUI automation, potentially accelerating adoption of AI-driven automation in enterprise software, mobile applications, and web services. The open-source release democratizes access to state-of-the-art GUI agent technology, enabling broader research and commercial applications. The platform-agnostic design—operating at the pixel level regardless of underlying OS or application type—suggests future convergence toward universal automation agents that can seamlessly handle heterogeneous digital environments. The multi-stage training approach combining massive mid-training with online reinforcement learning may become a standard paradigm for training long-horizon task agents. Support for 40+ mainstream Chinese applications indicates potential for localized AI agent ecosystems, with implications for regional technology sovereignty and alternative development pathways outside Western AI infrastructure. The demonstrated scalability across parameter sizes (2B-30B) suggests viable deployment options ranging from edge devices to cloud infrastructure, potentially enabling widespread integration into existing software stacks.