The 27B vision-language agent model is built on Qwen3.8-27B. Its code is available in the LBS branch of the UI-Venus repository.
The model is designed to select and zoom image regions while answering. The model card specifies up to two crops per turn and three crop calls per episode. inclusionAI says training used about 10,000 synthetic GUI-grounding SFT examples and 80 GRPO steps, with no general-domain VQA data.
With agentic zoom-in, the card reports 81.15% on ScreenSpot-Pro, 84.57% on OS-World-G (Refine), 59.67% on PerceptionBench, 89.13% on HR-Bench 4K, and 87.75% on HR-Bench 8K. PerceptionBench used a self-hosted DeepSeek judge. The source does not establish that these results are directly comparable with other models; it explicitly says chart comparisons are not controlled.