Release Notes

This page contains release notes for the Cosmos core models.

3.2

• Interactive World Simulation with Cosmos for Physical AI has been released. This includesexample model checkpoints and training recipes.

3.1

• Cosmos3-Edge–a 4B tier model for on-device, real-time deployment–has been released.

3.0

• Cosmos 3 merges separate models into a single suite of omnimodal world models, which are built on a unified Mixture-of-Transformers architecture.

2.2

• The Cosmos-Reason2 model has been added.

1.2

• The Cosmos-Reason1 model code base and examples have been restructured.

1.1

• Cosmos has been divided into three separate repositories: Predict1, Transfer1, and Reason1.

• The Predict1 Text2World and Video2World models now support multi-view generation.

• Inference and post-training scripts for Predict1/Transfer1 now operate using Conda environments.

1.0

Initial release

• Pre-trained diffusion world foundation model (WFM) with Text2World and Video2World generation

• Pre-trained autoregressive with Video2World generation