UGround-V1

UGround is a universal visual grounding model for GUI agents from the Ohio State NLP Group and Orby AI, published as an ICLR 2025 oral paper (arXiv 2410.05243). Its core contribution is scale: the team built what they describe as the largest GUI grounding dataset to date, covering 10 million GUI elements and their referring expressions across 1.3 million screenshots, then trained UGround on it so agents can locate interface elements purely visually.

The Qwen2-VL-based V1 series, in 2B, 7B and 72B sizes, was released in January 2025, with the team reporting state-of-the-art results on the ScreenSpot-Pro grounding benchmark. Training data, inference code and weights are all public. Check the model card for the exact license before commercial use.