World foundation model for physical AI; trained on 20 million hours of video with flow-matching architecture.