ChainThink reports that on April 22, according to an official announcement, the world model company Odyssey released Odyssey-2 Max.
This model has three times the parameters and ten times the training compute of its predecessor, Odyssey-2 Pro, making it the largest general-purpose world model the company has ever developed.
Unlike bidirectional video models such as Sora and Veo, Odyssey-2 Max employs a causal autoregressive architecture that predicts the next frame sequentially and accepts real-time interactive inputs, enabling continuous generation of simulated footage for over 120 seconds without crashing.
In terms of physical simulation accuracy, the Odyssey-2 Max scores 58.52 on the VBench 2 physical subtask, surpassing the Odyssey-2 Pro’s 49.67 and NVIDIA’s Cosmos-Predict2.5-14B’s 44.92.
Scored 93.02 on the PAI-Bench physical sub-item.
This model is based on the Autoregressive Diffusion Transformer (AR DiT) architecture, employs continuous flow matching for generation, and achieves real-time inference through few-step denoising distillation. Training was conducted in three stages across hundreds of NVIDIA B200 GPUs, including large-scale video pre-training, interaction and task conditioning, and long-sequence stability training.
Odyssey was founded by CEO Oliver Cameron and CTO Jeff Hawke, both from the autonomous driving industry, and has raised a total of $27 million in funding to date.
Odyssey-2 Max is now open to private beta partners in the fields of robotics, gaming, simulation, defense, and interactive systems.