World models
Odyssey-3 brings real-time physics to generated worlds
A 14-billion-parameter world model now turns a text prompt into an environment you can walk through, with objects that move and collide as you play. Here is what it changes for game makers.
On 8 October, Odyssey released Odyssey-3, a foundation world model that generates an interactive environment from a text prompt and computes how objects move and interact inside it in real time. A free browser preview lets anyone walk through a generated world in first or third person, steer the camera separately from movement and add events while the world is running.
The base model is reported at 14 billion parameters and renders at 832×480; the Pro version runs at 1280×720. Odyssey reports a score of 66.1 on the Physics-IQ Verified benchmark using best-of-eight sampling, the highest published result to date. Independent coverage measured 63.37 on a single-generation basis, which is the fairer number to quote.
Why it matters
The important change is not image quality, it is latency. By distilling an autoregressive diffusion model down to a few generation steps, Odyssey has turned a video generator into something closer to a simulator: you act, and the world answers immediately. That request-and-response loop is the basic unit of play.
Physics is the other half. Early world models produced convincing frames that fell apart when you pushed on them. A model that keeps objects behaving consistently is a model you can start to design rules around.
What it does not do yet
Resolution is still modest, the public preview is a research demo rather than a developer product, and Odyssey-3 API documentation and pricing are not yet published. Long-session memory, the thing that lets a world remember what you did an hour ago, remains the open problem across the field.
Our take
World models are converging on the layered approach we argue for in AI models as game generators: something must keep the truth of the world, and something must perform it. Real-time physics moves the performance layer forward fast. The studios that win will pair it with authored canon and persistent state.