Odyssey-2 Max

Odyssey-2 Max is Odyssey's largest and most powerful general-purpose world model, designed for real-time interactive simulation. It uses autoregressive next-state prediction to achieve high-fidelity physical accuracy, long-horizon stability, and open-ended interaction across evolving worlds. Unlike bidirectional video models (e.g., Sora, Veo, Kling, Runway), which generate fixed sequences, Odyssey-2 Max is causal, predicting each frame from prior states and user actions, enabling true real-time interactivity. Built for researchers and developers in robotics, gaming, defense, and healthcare, Odyssey-2 Max excels at simulating open-ended futures. It demonstrates that scaling next-state prediction yields emergent physics understanding—objects move, interact, and change according to learned dynamics. The model achieves state-of-the-art scores on VBench 2 Physics (58.52) and PAI-Bench Physics (93.02), while running in real time. Key architectural innovations include a proprietary KV cache enabling sequences 20x longer than prior work, causal attention with local and global context for temporal coherence, and flow matching in continuous latent space for high-fidelity visuals. The model underwent three-stage training: general visual dynamics, interaction conditioning, and long-horizon stability. Odyssey-2 Max is available now in private beta. What sets Odyssey-2 Max apart is its commitment to causal, autoregressive world modeling. At roughly 3x the size of Odyssey-2 Pro, it shows consistent improvements across motion smoothness, subject consistency, and image quality. The model generates over 120 seconds of continuous simulation, all conditioned on user actions, making it a powerful platform for interactive applications.