New World Model Odyssey-3 scores 66.1 in physics benchmark test, outperforms FLUX 3, now open for free trial.

Beating AI Express: World model company Odyssey has officially launched Odyssey-3, with its Flash version now available for free trial. The model can generate explorable worlds in real time from text, with visuals updating dynamically as users move or rotate the camera. The team also showcased experiments applying the same model to robotic arms, humanoid robots, vehicles, and game characters. Odyssey-3 first learns object movement and action outcomes from internet videos, game play recordings, and physical simulations. When deployed for robots or vehicles, the main pre-trained model parameters remain unchanged; the team only needs to train additional small control modules to translate the model’s learned knowledge into operational instructions. While different machines still require separate adaptation, full retraining of the entire model is unnecessary. Odyssey trained a dedicated control module using just 20 hours of simulated driving data, enabling a car to operate on real roads in India. In tests, the simulated data training scheme delivered approximately 77% of the driving distance between two safety officer takeovers compared to the real driving video training scheme. For the humanoid robot developed in partnership with Flexion, after training on tens of hours of remote control demonstration data, it can complete tasks even under varying lighting conditions. In game experiments, the control module learned from roughly 2 hours of Grand Theft Auto V (GTA V) gameplay footage, and without additional training, it enabled characters to ride horses and move in Red Dead Redemption 2. In the Physics-IQ Verified physical video prediction test, Odyssey-3 Pro scored 66.1 points, surpassing FLUX 3 [large]’s 64.35 points, with both results being the best from 8 trials. Odyssey also released WorldMark evaluation results, where the model secured first place in three out of four interaction scenarios.