Odyssey-3 is a free generative world model you can try right now
California startup Odyssey releases its world model to the public, letting anyone build interactive environments from text.
A world you can walk into
Odyssey, a California-based AI company, has opened a free public research preview of its new Odyssey-3 world model 1. Instead of just generating images or videos, the model creates entire interactive environments that you can explore in real time from a text description.
You type something like "a rainy cyberpunk street" and the model builds a 3D-like world you can walk through, switch between first-person and third-person views, trigger events, and watch the environment respond as if it were truly alive.
How it works
Odyssey-3 uses what is called an autoregressive diffusion transformer — a kind of AI that predicts the next piece of data in a sequence while also filling in missing details, like guessing the next frame of a video based on what came before 1. It was trained on internet videos with event descriptions, video game footage paired with keyboard and mouse inputs, and simulated physics interactions.
The base model has 14 billion parameters and generates video at 832 by 480 pixels, while the more powerful Odyssey-3 Pro version supports 1280 by 720 1.
Physics benchmarks with a catch
The company says Odyssey-3 Pro scores 66.1 on Physics-IQ Verified, a benchmark that tests how well AI models understand real-world physics like fluids, light, and motion 1. But there is a catch — that top score came from a single test run where the best of eight generated videos was picked for each task, rather than reporting results across four runs as the rules require . The average across four runs was 63.37 .
On WorldMark, another benchmark that tests consistency and control in generated worlds, Odyssey-3 ranked first in three of four categories based on the company's own evaluations 1.
Beyond games: robots and drones
Odyssey wants the same model to power many different systems at once. It has used Odyssey-3 to control robotic arms, train drone controllers, and even make AI agents play GTA V autonomously 1. A controller trained on about two hours of game footage could also transfer its skills to Red Dead Redemption 2 without extra training .
The company is also working with Swiss robotics firm Flexion on humanoid robot controllers that reportedly perform more reliably than comparison models 1.
Odyssey was founded in 2023 by Oliver Cameron and Jeff Hawke, who previously spent ten years building driverless cars 3. In June 2026 the company raised $310 million, with investors including Amazon and AMD Ventures 1.
Key facts
- Odyssey-3 is a 14-billion-parameter generative world model available as a free public preview
- It generates interactive 3D-like environments from text prompts in real time
- Physics-IQ Verified score of 66.1 has caveats — it used a selection method not allowed by benchmark rules
- The model can control robots, drones, and autonomous game characters
- Odyssey raised $310 million in June 2026 from investors including Amazon and AMD Ventures
Sources
- Odyssey-3 is a new generative world model that you can try for free — the-decoder.com
- Odyssey-3 Launches as Free Generative World Model Preview — techaimag.com
- Meet Odyssey-3: Our Most Powerful Foundation World Model
- Introducing Odyssey-3: A General-Purpose Physical Intelligence
- Odyssey Launches Odyssey-3 World Model Research Preview for Free — gate.com