Odyssey-3 has left the lab behind. A free research preview of the world model is now open from the California company, letting a text prompt become an interactive environment that people walk around inside, live.
The model got its first public showing on September 15 from founders Oliver Cameron and Jeff Hawke, pointed at games, autonomous driving and robotics. Access, technical disclosure and benchmark results are the additions.
Odyssey-3 Flash powers the free demo. A visitor picks first-person or third-person, moves through the scene, sets off events, and watches the system answer back with no rendering lag.
Beneath that sits an autoregressive diffusion transformer, drawing each fresh frame from the ones before it plus the user’s latest action, which gives the model a continuous grip on cause and effect.
Internet video carrying event captions went into training, along with game footage matched to keyboard and mouse input and simulated physical interactions. A second technique trims the compute required per step, which is what makes real-time generation workable.
At its base the system holds 14 billion parameters and paints 832 by 480 pixels; a Pro tier climbs to 1280 by 720.
Odyssey reports that Pro hit 66.1 on the Physics-IQ Verified video benchmark, a test of how faithfully a model extends real experiments across fluid mechanics, optics, solid mechanics, magnetism and thermodynamics. That number, though, came from one run in which a selection method chose the best of eight outputs, a shortcut the benchmark’s own rules bar for record claims. Four standard runs left Pro at an average of 63.37. The leaderboard carries both figures.
API access is open to developers who apply.
