Odyssey-2 Max

Odyssey‑2 Max is a general‑purpose world model that predicts future visual states from prior observations and actions, enabling real‑time, continuous simulation of physical dynamics. It is trained directly on visual recordings of real‑world activity, allowing it to internalize object motion, interaction, and biomechanical processes without relying on textual descriptions. The model operates autoregressively, generating each frame conditioned on earlier frames and user actions, which supports open‑ended rollouts that evolve interactively.

The system is positioned for researchers and developers who need high‑fidelity physical simulation for domains such as robotics, scientific modeling, gaming, defense, and healthcare. Its causal architecture distinguishes it from bidirectional video generators that produce fixed‑prompt outputs, providing a framework for interactive, step‑by‑step prediction.

Odyssey‑2 Max claims the highest physics benchmark scores among evaluated world models while maintaining real‑time performance, reflecting its focus on physical accuracy as the model scales. It is currently available in private beta and is described as an experimental, large‑scale advancement in multimodal world simulation.