IT之家 · 智能时代 · 10/9/2026, 14:41:48
Odyssey-3 Launches Foundation World Model: Autoregressive Diffusion Tops Physics-IQ with Real-Time Causal Prediction
Odyssey has released the Odyssey-3 series of foundation world models, with the Pro version setting a new record score of 66.1 on the Physics-IQ Verified benchmark. Utilizing an autoregressive diffusion transformer architecture, the model learns physical dynamics and causal relationships from visual observations, enabling real-time environment prediction based on user actions like viewpoint shifts or event triggers, while ranking first across multiple metrics in WorldMark evaluations.
SOURCE COVERAGEOriginal coverage
On October 9, Odyssey announced the launch of its Odyssey-3 series foundation world models. Notably, the Odyssey-3 Pro achieved a score of 66.1 on the Physics-IQ Verified video-to-video benchmark, setting a new record for the leaderboard.
Note from IT Home: A Foundation World Model is a general-purpose AI model that learns physics, dynamics, and causal relationships from extensive visual observations. It can predict how objects move and interact, as well as how contexts evolve over time, enabling it to simulate environments or train control policies for systems such as robots and autonomous vehicles.
The Physics-IQ Verified benchmark was jointly developed by Anates Labs and DeepMind. It requires models to continue real-world physical experiment videos and compares predicted outcomes against actual results, covering domains such as fluid dynamics, optics, solid mechanics, magnetism, and thermodynamics.
Regarding the models, the Odyssey-3 series includes Standard and Pro versions. The Standard version balances physical accuracy with generation costs, while the Pro version offers stronger physical prediction capabilities.
In terms of architecture, Odyssey-3 employs an autoregressive diffusion transformer architecture capable of generating embodied environments from prompts. As users shift perspectives, take actions, or introduce events during generation, the model combines existing observations with the latest inputs to predict environmental changes in real time.
This preview release supports both first-person and third-person navigation, as well as independent camera movement. Users can navigate within the environment or trigger events to observe how the model responds to changes in environmental state.
Performance-wise, in the WorldMark evaluation, Odyssey-3 ranked first across three categories: first-person stylized, third-person realistic, and third-person stylized environments, achieving scores of 77.2, 79.0, and 76.3, respectively.