Odyssey, a California-based AI company, has opened a public research preview of its world model Odyssey-3. Anyone can type a text description, get a 3D environment and move through it in real time, choosing first- or third-person view and triggering events the model handles instantly. The free demo runs on the Odyssey-3 Flash variant; developers can apply for API access.
The model is built on an autoregressive diffusion transformer that generates each new frame from previous frames and user actions. According to the company, it learned physical relationships and cause and effect from internet videos, video game footage paired with keyboard and mouse inputs, and simulated physical interactions. An extra training technique reduces the number of compute steps required, which makes real-time generation possible.
Specs and benchmarks
According to the source, the base model has 14 billion parameters and generates video at 832 × 480 pixels; Odyssey-3 Pro supports 1280 × 720. On the Physics-IQ Verified benchmark, which tests fluid mechanics, optics, solid mechanics, magnetism and thermodynamics, Odyssey-3 Pro scores 66.1 points on the video-to-video test. That figure needs a caveat: it comes from a single run in which a selection method picked one of eight generated videos per task, while the benchmark's rules require four runs with standard deviation for any record claim. Without that selection, the four-run average is 63.37 points. Both results appear on the benchmark page but were submitted by Odyssey itself.
On the WorldMark benchmark, which measures control-instruction following, image quality and consistency over time, Odyssey-3 ranks first in three of four categories: 77.2 for first-person stylized, 79.0 for third-person real and 76.3 for third-person stylized. In first-person real it places third with 80.6 points.
One model for robots, drones and games
Odyssey wants the same world model to handle very different tasks: robotic arms, drones and game characters. For each use, the model is paired with a specialised controller that turns its predictions into concrete commands. According to the company, an AI built on Odyssey-3 controlled several robotic arms after a few dozen hours of demonstration data, and the robots recovered on their own from failed grasps, situations absent from the training data. For humanoid robots, Odyssey works with Swiss firm Flexion, whose controllers built on Odyssey-3 are said to be more reliable than comparison models even when conditions change. A controller trained on simulated flight data let a drone avoid obstacles and reach targets in a virtual indoor environment.
In games, Odyssey showed an AI autonomously driving and fighting in GTA V. A controller trained on about two hours of GTA footage transferred its skills to Red Dead Redemption 2 without extra training, moving a character on horseback. Odyssey also plans to use the model to train AI agents that receive tasks in natural language and try to complete them through their own actions in a generated environment.
For now, the public preview covers only interactive environment generation: the robotics, drone and game-controller applications remain separate research work, not yet available.
The race for world models
Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke and in June 2026 raised $310 million from investors including Amazon and AMD Ventures. World models aim to give AI systems an understanding of spatial and physical relationships. Google DeepMind pursues a similar approach with Genie 3, while Fei-Fei Li's startup World Labs works on related technology. Also in this space is GPT-6 Astra, which turns real videos into interactive 3D worlds, while Tencent's WorldClaw focuses on explicit, editable 3D worlds that separate terrain and objects.
No comments yet. Be the first!