Odyssey opens preview of its 14-billion-parameter Odyssey-3 world model

Odyssey has launched a public research preview of its generative world model, Odyssey-3. The system generates interactive environments directly in the browser, runs in real time, and delivers top-tier results on physics simulation benchmarks.

Architecture and training

The model is built on an autoregressive diffusion transformer architecture, which generates new video frames sequentially based on previous frames and user actions. The system was trained to recognize physical patterns and cause-and-effect relationships from visual observations.

The base version has 14 billion parameters and generates video at 832×480 pixel resolution, while the Pro version supports 1280×720. The system was trained on three types of data: internet videos with event descriptions, video game footage paired with corresponding keyboard-and-mouse inputs, and simulated physical interactions. A specialized training method reduced the number of computation steps required, enabling real-time generation.

Users can generate interactive environments from text prompts and choose between first-person and third-person perspectives. Within the generated worlds, they can move around, trigger events, and watch the model respond in real time.

Benchmark results

On the Video-to-Video benchmark, which evaluates behavior in fluid dynamics, optics, rigid-body mechanics, magnetism, and thermodynamics, Odyssey-3 Pro scored 66.1 points using a single run with best-of-eight selection. This result does not meet the official record standard, which requires four runs with a reported standard deviation. Without the selection method, the average score was 63.37 points.

On the WorldMark benchmark, which assesses instruction-following, visual quality, and world consistency, Odyssey-3 leads in three of four categories: • First-Person Stylized: 77.2 points • Third-Person Real: 79.0 points • Third-Person Stylized: 76.3 points • First-Person Real: 80.6 points (third place)

From robotics to video games

Odyssey is working toward a universal world model for a range of tasks — from controlling robot arms to drone navigation and video game character control. In each case, the model is paired with a specialized controller that translates its predictions into concrete commands.

In tests, an AI trained on Odyssey-3 controlled several robot arms. Just a few dozen hours of demonstration video were enough to train it, and the robots were able to recover from failed grasps even though such scenarios weren't part of the training dataset.

Odyssey is partnering with the Swiss company Flexion Robotics, which has developed controllers for humanoid robots based on Odyssey-3. These controllers perform more reliably than competing models, even when conditions change.

For drones, Odyssey trained a controller on simulated flight data. The drone was able to avoid obstacles and fly to target positions on command within a virtual environment.

In the video game domain, the system demonstrates autonomous control in GTA V, including vehicle operation and combat. A controller trained on roughly two hours of GTA V footage successfully transferred its skills to Red Dead Redemption 2 without any additional training, controlling a character on horseback.

The company plans to use the world model to train AI agents. In a demonstration, an AI agent received a task in natural language and carried it out through its own actions within an environment generated by Odyssey-3 — establishing a new training paradigm in which agents learn from the outcomes of their own actions.

Investment and competition

Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke. In June 2026, the company secured investment from Amazon and AMD Ventures. In the world model sector, competitors include Google DeepMind with its own approach to interactive worlds, and World Labs (founded by Fei-Fei Li), which AMD announced its intent to acquire for $8.2 billion in late September 2026.