Odyssey-3
#5 in 3D & World GenerationUnknown · v3 · 2× · last seen Sep 19, 2026
35
Momentum
Odyssey-3 is a fundamental "World Model" developed by Odyssey, designed as an autoregressive diffusion transformer and trained on a broad collection of visual observations of the physical world. Rather than training a separate model for each machine, a frozen backbone serves as a shared foundation that is adapted to robotic arms, humanoids, vehicles, drones, and video game characters through small, task-specific Action Decoders. The model was unveiled on September 15, 2026 by Odyssey founders Oliver Cameron and Jeff Hawke; public access and pricing and licensing details had not been published at the time of announcement.
Momentum trend
21.06.19.09.
Features
| Fine-Tuning | Adapted to new systems via a lightweight 'action decoder' trained with only tens of hours of robot demonstration data |
| License | Not specified (format, license, checkpoint access and API limits not yet disclosed by the company) |
| Platform | Controls robot arms, humanoids, cars, drones and video-game characters via one shared frozen backbone |
| Price | Not announced – no public download or pricing at time of announcement |
| Release Date | Unveiled September 15, 2026; public release planned in coming weeks |