UBTECH published Thinker-WM on Hugging Face on 29 Sep 2026, with a code repository created the same day. The model card describes it as “a unified multimodal world model that jointly models and generates RGB, depth, and action.” It was pretrained on 20,000 hours of egocentric data and 10,000 hours of robot data and is fine-tuned from Wan2.2-TI2V-5B.
The weights carry a Creative Commons BY-NC-SA 4.0 license, which rules out commercial use. The model card reports first place on three public benchmarks: an overall score of 4.66 on EWMBench, 0.739 on the WorldOlympiad robotics leaderboard and an EWMScore of 71.73 on WorldArena Track-1. It also reports a 75.0% closed-loop success rate on the RoboTwin simulation benchmark.
All results are reported by UBTECH. The model card names no UBTECH robot as the target hardware, so the item is linked to the company card only.