| Cosmos World Foundation Model Family | |||
|---|---|---|---|
| Cosmos-Predict1 | Code | Models | Paper |
| Cosmos-Transfer1 | Code | Models | Paper |
| Cosmos-Reason1 | Coming soon | Coming soon | Paper |
NVIDIA Cosmos is a developer-first world foundation model platform designed to help Physical AI developers build their Physical AI systems better and faster. Cosmos contains
- Pre-trained models (available via Hugging Face) under the NVIDIA Open Model License that allows free commercial use.
- Pre-training, post-training, and inference code (available in native PyTorch) under the Apache 2 License.
There are three main model families in Cosmos World Foundation Model Platform.
-
Cosmos Predict: a collection of general-purpose world models for future state prediction.
-
Cosmos Transfer: a collection of multimodal conditional world generation model for various domain transfer applications such as Sim2Real.
-
Cosmos Reason: a collection of Physical AI reasoning models for planning and critics.
Being a minimalist, we have these individual models in individual repositories under nvidia-github.
predict1_text2world.mp4
predict1_video2world.mp4
Cosmos-Transfer LiDAR + HDMap Conditional Inputs -> World
transfer1_lidarhdmap.mp4
Cosmos-Transfer Multimodal Conditional Inputs -> World
transfer1_multimodal.mp4
Cosmos-Reason Physical AI Planning
reason1_example1.mp4
| Paper Title | Summary | Authors | Date |
|---|---|---|---|
| Cosmos-Reason1: From Physical Common Sense To Embodied Reasoning | Introduces a reasoning model for physical AI that combines common sense knowledge with embodied reasoning capabilities. | NVIDIA | 2025-03-19 |
| Cosmos-Transfer1: Conditional World Generation with Adaptive Multimodal Control | Presents a multimodal model for conditional world generation with adaptive control mechanisms. | NVIDIA | 2025-03-18 |
| Cosmos World Foundation Model Platform for Physical AI | Overview of the Cosmos platform, its architecture, and applications in physical AI systems. Introduction of Cosmos-Predict1 world models. | NVIDIA | 2025-01-06 |
For native PyTorch developers, we provide native PyTorch training and inference scripts in nvidia-github. For Nemo developers, please refer to README_CES2025.md.