Odyssey has released the Odyssey-2 Max world model, with its physics simulation scoring ahead of similar models.

According to Dynamic Beating monitoring, the world modeling company Odyssey has released Odyssey-2 Max, with a model size three times that of its predecessor Odyssey-2 Pro and a training computational load ten times higher, making it the company's largest universal world model to date. Unlike bidirectional video models like Sora and Veo, Odyssey-2 Max adopts a causal autoregressive architecture to predict the next state frame by frame and accept real-time interactive input, capable of continuously generating simulated scenes of over 120 seconds without crashing.

In terms of physical simulation accuracy, Odyssey-2 Max scored 58.52 in the VBench 2 physics subtask, higher than Odyssey-2 Pro's 49.67 and NVIDIA Cosmos-Predict2.5-14B's 44.92; with a score of 93.02 in the PAI-Bench physics subtask. The model is based on the Autoregressive Diffusion Transformer (AR DiT) architecture, uses continuous flow matching generation, and achieves real-time inference through few-step denoising distillation. Training was carried out on hundreds of NVIDIA B200 GPUs in three stages: large-scale video pretraining, interaction and task conditioning, and long sequence stability training.

Odyssey, founded by Oliver Cameron (CEO) and Jeff Hawke (CTO) from the autonomous driving industry, has previously raised a total of $27 million. Odyssey-2 Max is now available for Private Beta to partners in the robotics, gaming, simulation, defense, and interactive systems sectors.