WoW-1 (Wu 1.0)

WoW-1 — World-Omniscient World model, branded by its makers as Wu 1.0 — is an open-weights generative world model from the Beijing Innovation Center of Humanoid Robotics (X-Humanoid), the state-backed centre behind the Tiangong humanoids. The first checkpoints appeared on 15 October 2025, with a small 1.3-billion-parameter build following on 28 October. A world model predicts what a scene will look like after an action, which is how a robot can rehearse a movement before making it. The premise of WoW is a criticism of how video models are usually built: systems such as Sora learn by passively watching video, and the authors argue that physical causality cannot be picked up that way. So the 14-billion-parameter main build was trained on 2 million real robot interaction trajectories — recordings of machines actually pushing, grasping and dropping things. The centre states the pretraining mix as 15% robot video and 85% its RoboMIND dataset, and the model is built on Alibaba's Wan video generator. The honest part of the research report is the failure mode it documents. The model's grasp of physics turns out to be a probability distribution over plausible outcomes, which in practice means instability and what the authors call physical hallucination — objects behaving in ways they could not. Two additions try to contain it. SOPHIA is a vision-language critic that judges the generated video and rewrites the instruction until the result looks physically sound. A separately co-trained inverse dynamics model then converts the accepted video into executable robot actions, closing the loop from imagination to movement. The team also published WoWBench, its own benchmark for physical consistency and causal reasoning in video, and reports state-of-the-art results on it in both human and automated scoring — the usual caveat applies, since the benchmark and the model come from the same authors. The weights are on Hugging Face under the MIT licence in several sizes: 14B trained on the full 2 million trajectories, a 14B checkpoint at 600,000, DiT builds at 2B and 7B, and a 1.3B build for smaller machines. Not every repository in the set carries a licence file, so anyone reusing the smaller DiT checkpoints should check the terms before shipping them.

▸Videos

No videos yet.

▸Reviews