Odyssey has opened a public research preview of Odyssey-3, a 14-billion-parameter world model that generates interactive environments from text prompts in real time. Users can move through scenes, change perspective and trigger events in a free demo running the faster Flash version; developers can apply for API access.
The autoregressive diffusion transformer predicts new video frames from earlier frames and user actions. Odyssey says training combined described internet video, game footage paired with keyboard and mouse inputs, and simulated physical interactions. The base model outputs 832 by 480 pixels, while Odyssey-3 Pro supports 1280 by 720.
The company reports uses beyond scene generation through specialized controllers for robot arms, simulated drones and game characters. Those demonstrations are company results, not broad proof that one controller transfers reliably across real machines and environments.
Its strongest Physics-IQ Verified score also needs qualification. A 66.1 result selected one of eight generations per task from a single run, while record rules require four runs and a reported standard deviation. Without that selection method, Pro averaged 63.37 across four runs. The public preview makes the model testable, but independent evaluation remains necessary for its physics and control claims.