Odyssey launches Odyssey-3 foundation world model on October 10 with Pro scoring 66.10 on Physics-IQ Verified

Odyssey launched Odyssey-3 foundation world models on October 10, including a standard version and a Pro version. The standard Odyssey-3 is a 14B parameter auto-regressive video model at native 480p; Pro runs at native 720p and shares the same foundation weight family. On the Physics-IQ Verified video-to-video leaderboard, developed with DeepMind and Anates Labs, Pro scores 66.10 with best-of-eight sampling. Standard single-shot generation scores 61.56 and best-of-eight sampling reaches 64.43, while Pro single-shot is 63.37. Odyssey-3 uses an auto-regressive diffusion transformer architecture: a user enters a prompt, the model generates an embodied environment, and during generation the user can move the viewpoint, take actions, or introduce events while the model continuously predicts how the environment evolves. On the WorldMark benchmark, Odyssey-3 ranks first in all three categories, scoring 77.2 in first-person stylized, 79.0 in third-person realistic, and 76.3 in third-person stylized. Odyssey also shows demos with robot arms, Flexion humanoid robots, and a driving policy trained on roughly 20 hours of driving data controlling a vehicle on roads in India; the drone experiments remain simulated. The Flash variant is open in a browser as a research preview.