UI-Venus-2-9B is a general-purpose foundation GUI agent from Ant Group's inclusionAI, initialized from Qwen3.5-9B and designed to operate across mobile applications, web platforms, and desktop operating systems through a unified closed-loop reasoning-action framework that observes the interface, reasons about task state, executes an action, and incorporates environmental feedback into its next decision. It was trained via a three-stage pipeline — multimodal mid-training on large-scale simulated trajectories, offline RL developing domain-specialized policies across Grounding/CAPTCHA/Mobile/Web/Computer task families, and multi-teacher on-policy distillation consolidating those specialists into one unified policy — supported by a deep-research-driven task-generation pipeline spanning 170+ multilingual mobile apps, a 4,000+ domain web pool across 19 categories, and native desktop environments, with keypoint-grounded, multi-model-voting verification designed to resist reward hacking. It achieves near state-of-the-art results for its scale across mobile use (80.2 on AndroidWorld, 65.8/75.2 on MobileWorld), computer use (70.8 on OSWorld-Verified, 48.0 on DeskCraft), web navigation (90.8 on WebVoyager, 74.0 on Online-Mind2Web), GUI grounding (73.0 on ScreenSpot-Pro), and CAPTCHA solving (78.1 on VenusBench-CAPTCHA), while also showing markedly improved consequential-action safety over its Qwen3.5-9B base — reducing Attack Success Rate from 25.3% to 11.3% on OSHarm and from 79.4% to 48.8% on OSBlind. It's served via vLLM with an OpenAI-compatible endpoint (requiring the separate UI-Venus code repository for action parsers and closed-loop execution), and while released with full-parameter weights, its final license is still pending confirmation at time of release; deployments are still recommended to retain human confirmation around consequential operations.
