California-based AI company Odyssey launched a public research preview of its generative world model, Odyssey-3, allowing users to explore interactive environments generated from text prompts in real time. Founders Oliver Cameron and Jeff Hawke first unveiled the model on September 15, and it is now publicly accessible via a free online demo running on Odyssey-3 Flash, while developers can also apply for API access.
Built on an autoregressive diffusion transformer, Odyssey-3 features 14 billion parameters in its base version and is designed to serve as a foundation for controlling various systems, such as robotic arms, drones, and video game characters. The model was trained on internet videos, video game footage, and simulated physical interactions, utilizing an extra training technique that reduces compute steps to enable real-time video generation.
In benchmark evaluations, the more powerful Odyssey-3 Pro achieved notable scores on physics and world-modeling tests, though some results involved selective video choices rather than standard multi-run averages. Odyssey aims to deploy the same versatile model across multiple domains by pairing it with specialized controllers to translate predictions into concrete operational commands. This release matters because it provides public access to an advanced real-time world simulator that bridges the gap between text prompts and interactive physical environments for robotics and gaming.