On 6 November 2026, Hugging Face reports that the Nvidia Cosmos world model development platform has added four new open-weight checkpoints to its Cosmos 3 model family. These additions focus on generating interactive world simulation and multi-sensor multi-view driving scenes for research and development.
Key takeaways
- Nvidia introduced four new open-weight checkpoints to the Cosmos 3 family on 6 November 2026.
- The models support tasks like synthetic driving data generation and bimanual robotics simulation.
- Each of the four checkpoints contains 16 billion trainable parameters.
- Commercial and non-commercial use is permitted under OpenMDW1.1 terms.
What is the Nvidia Cosmos 3 interactive world simulation platform?
The updated Nvidia platform provides tools to build responsive simulated environments. Robotics needs systems that can handle changing scenes and viewpoints, and these models offer ways to meet that demand. According to Hugging Face, the new checkpoints include Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer, and Cosmos3-Nano-Sim-Bimanual.
Each model serves a specific purpose in testing and training autonomous agents. For instance, Transfer-Auto creates multi-view driving data with varied weather and lighting, while Sim-Auto streams synchronized camera and LiDAR observations. Sim-Transfer turns simulator controls and text prompts into photorealistic video, and Sim-Bimanual predicts video from robot or camera actions.
How do open weights and autonomous driving simulation fit together?
For autonomous driving, appearance and geometry must agree completely. Synchronized camera and LiDAR generation connects visual observations directly with 3D sensing. Scene and traffic-signal controls allow researchers to alter road layouts and signal states to study policy responses.
All four models are post-trained examples of Cosmos 3 Nano, and all feature 16 billion trainable parameters. They use OpenMDW1.1, which permits both commercial and non-commercial use under its specific terms. Nvidia intends to give researchers and developers a shared starting point for adapting models and comparing methods.
Why it matters
Developers and researchers working on physical AI systems gain access to standardized starting points for custom sensor rigs and robot embodiments. Because the work ahead involves measuring where models faithfully support policy learning and understanding their limits, these open checkpoints provide a practical foundation for community contributions and further improvements.
Frequently asked questions
The four new models are Cosmos3-Nano-Transfer-Auto, Cosmos3-Nano-Sim-Auto, Cosmos3-Nano-Sim-Transfer, and Cosmos3-Nano-Sim-Bimanual.
All four new models are post-trained examples of Cosmos 3 Nano and each feature 16 billion trainable parameters.
The models use OpenMDW1.1, which permits both commercial and non-commercial use under its terms.
It is a bidirectional model for offline synthetic driving data generation that creates synchronized video across up to 11 camera views and supports joint RGB-LiDAR generation.
It simulates how an environment evolves in response to a bimanual robot's actions, supporting interactive two-arm simulation for YAM and AgiBot embodiments.
Based on reporting by Hugging Face. Read the original article
