Odyssey-3 goes public: a free world model you can steer in real time

Odyssey opened a public research preview of Odyssey-3 on Sunday, and anyone can try it: type a sentence, step inside the world it generates, and watch it respond in real time. The same 14-billion-parameter model already steers robot arms, drones, and a car on real roads in India.

Odyssey opened a public research preview of Odyssey-3 on Sunday, and anyone can try it: type a sentence, step inside the world it generates, and watch it respond in real time. The same model already steers robot arms, drones, and a car driving on real roads in India — from one shared foundation, trained once on how the world moves.

What you actually get

The free online demo runs on Odyssey-3 Flash. Describe an environment in text, and the model generates it frame by frame — in real time, not after a render queue. You choose a first-person or third-person view, move through the scene, trigger events, and the world updates as you act, according to coverage of the launch.

That immediacy is the whole trick. Odyssey-3 is built on an autoregressive diffusion transformer: it predicts each new video frame from the frames before it and from your actions. A training technique that cuts the number of required compute steps makes the generation fast enough to be interactive — a world that dreams itself forward as fast as you can walk through it.

The base model carries 14 billion parameters and outputs video at 832 × 480 pixels; a heavier Odyssey-3 Pro runs at 1280 × 720, per Odyssey’s official benchmark submission. Developers can also apply for API access, so this preview is a recruiting poster as much as a product: Odyssey wants physical-AI builders tinkering with it.

The benchmark claims, with the caveats they deserve

Odyssey-3 Pro scores 66.1 points on the Physics-IQ Verified video-to-video benchmark, which Odyssey says is the highest reported score. The benchmark tests physical phenomena — fluid dynamics, optics, solid mechanics, magnetism, thermodynamics — by asking models to continue videos of real-world experiments. Here is the catch, and credit to The Decoder for spelling it out: that 66.1 comes from a single run where a selection method picked the best of eight generated videos per task. The benchmark’s own rules require four runs with a reported standard deviation for any record claim. Without the cherry-picking, Odyssey-3 Pro averaged 63.37 points across four runs. Both numbers are on the leaderboard, submitted by Odyssey itself.

On the WorldMark benchmark — which tests how well models follow control instructions, how good the visuals look, and whether simulated worlds stay consistent over time — Odyssey-3 ranks first in three of four categories on the company’s own evaluation: First-Person Stylized (77.2), Third-Person Real (79.0), and Third-Person Stylized (76.3). In First-Person Real it places third at 80.6. Take company-run evaluations for what they are; independent testing is now possible precisely because the model is public.

One model, many bodies

The pitch is not the demo — it’s the reuse. One model is pretrained on visual observations of the world, and small task-specific controllers translate its predictions into commands for different machines. In Odyssey’s tests, AI built on Odyssey-3 controlled multiple robot arms after only a few dozen hours of demonstration data, and recovered from failed grasps on its own even though those situations were never in the training data. Swiss robotics company Flexion built humanoid control policies on it that Odyssey says beat the tested VLA baselines when conditions changed — including lighting shifts that made the baselines fail.

The wilder demonstrations: a drone controller trained on simulated flight data navigating an indoor environment and dodging obstacles; a driving policy trained on roughly 20 hours of data driving a real car on roads in India with the Odyssey-3 backbone frozen; and an AI playing Grand Theft Auto V, steering vehicles and fighting enemies — where a controller trained on about two hours of GTA footage transferred to Red Dead Redemption 2 with no additional training, moving a character on horseback. Odyssey also showed AI agents pursuing natural-language goals inside generated environments, a potential new way to train agents by letting them learn from consequences rather than datasets.

Why this matters

The race to build world models is now the most expensive argument in AI. Google DeepMind is pursuing the same interactive-worlds approach with Genie 3; Fei-Fei Li’s World Labs is developing similar technology, and AMD announced in late September that it plans to acquire World Labs for roughly $8.2 billion. Odyssey’s counter-move is to make its world model the one you can actually touch. A playable model recruits the army of builders that a press release never will — the question now is whether the physics holds up once thousands of strangers start throwing rocks into it.