California-based company Odyssey has opened a public research preview of its world model Odyssey-3. The system generates interactive environments from text descriptions and lets users explore and influence them in real time, choosing first- or third-person views and independently controlling the camera.
Unlike conventional video generators that produce a finished clip, Odyssey-3 acts as a learned dynamical system. It continuously predicts how the scene evolves based on prior observations and the actions of a user or agent. The free demo runs on Odyssey-3 Flash; the Pro variant generates at higher resolution (1280 × 720).
How it works and what it can do
The model is an autoregressive diffusion transformer. It was trained on a mix of internet videos with event annotations, gameplay recordings with synchronized keyboard and mouse inputs, and simulated physical interactions. An additional distillation step reduces the compute steps needed for real-time generation.
According to Odyssey, the same base model can be adapted via specialized policies to control robotic arms, humanoids, vehicles, drones and even game characters. In demos, a controller trained on roughly two hours of GTA V footage transferred skills to Red Dead Redemption 2 without further training. The company also showed policies for Flexion robots and real-road driving in India with about 20 hours of data.
Benchmark results and caveats
Odyssey claims Odyssey-3 Pro scored 66.1 on the video-to-video track of Physics-IQ Verified, the highest reported figure. The test evaluates continuation of real physical experiments in areas such as fluid mechanics, optics and solid mechanics. On WorldMark, the model ranked first in three of four categories in the company’s own evaluation.
These numbers need context. The 66.1 score came from a best-of-8 selection in a single run; the average of four standard runs was 63.37. Results were submitted by Odyssey to the leaderboard. The base model has 14 billion parameters and generates at 832 × 480.
Why it matters
World models aim to give AI systems a causal understanding of physics and dynamics rather than just visual appearance. Making the model freely testable in the browser brings the approach within reach of more people, including game developers and robotics researchers. Odyssey, founded by Oliver Cameron and Jeff Hawke (with autonomous-vehicle backgrounds), raised $310 million in June 2026 from investors including Amazon and AMD Ventures.
The field remains competitive. Google DeepMind is pursuing similar lines with Genie, and World Labs (founded by Fei-Fei Li) was the target of an AMD acquisition announcement. The public preview of Odyssey-3 offers a concrete point for observing progress in AI-generated interactive environments.
Sources
- Official Meet Odyssey-3 page
- Free Odyssey-3 Flash demo
- The Decoder coverage
- Official X account (@odysseyml)
Transparency: This content was created, edited or reviewed with the assistance of artificial intelligence. Information was cross-checked against public posts on X and sources available on the internet. Consult the original sources for full context.
Por GeekikiBot