Odyssey releases world model Odyssey-3

Startup Odyssey, founded by former self-driving car developers, released the foundational world model Odyssey-3 on October 8, 2026, which generates an interactive environment from a text prompt and calculates object movement and interaction in real time.

Compiled by Sergey KostenchukPublished 2026-10-09Updated 2026-10-10

2026-10-09 Research NVIDIA

🤖 Odyssey releases world model Odyssey-3

The self-driving car creators' startup released a foundational world model on October 8: an autoregressive diffusion transformer builds an interactive environment from text and calculates object movement and interaction in real time. Odyssey-3 (832×480) and Pro (1280×720) versions are distilled to a few generation steps.

🌍 A universal simulator instead of private ones like CARLA: on top of a frozen world model, they train robotic arm decoders, Flexion humanoid policies, and real car driving in India on 20 hours of data. On Physics-IQ Verified — 66.1 points, ahead of NVIDIA Cosmos3, Veo 3.1, and Sora 2.

👤 Interactive mode is open for free in the browser: a world from text, first- or third-person walk, a separate camera, on-the-fly events. For physical AI developers — API via developer.odyssey.ml.

Source 1: https://odyssey.systems/meet-odyssey-3 Source 2: https://experience.odyssey.systems/

Sources