In brief
- According to NVIDIA’s announcement, each of the four Cosmos 3 Nano models has 16 billion trainable parameters.
- Transfer-Auto can generate synchronized video from up to 11 camera views and RGB–LiDAR data, according to the company.
- Sim-Auto generates sequential camera video and LiDAR observations from driving controls, NVIDIA says.
- The models are available for commercial and non-commercial use under the terms of OpenMDW1.1.
On 5 November 2026, NVIDIA added four open-weight checkpoints to the Cosmos 3 model family for interactive world simulation and multisensor driving scenes. The announcement was published on Hugging Face. According to NVIDIA, each model in the Cosmos3-Nano family has 16 billion trainable parameters, with applications ranging from driving data generation to predicting the visual outcomes of robot movements.
What are the four models designed to do?
Cosmos3-Nano-Transfer-Auto was designed to generate driving scenarios in different weather, lighting and time-of-day conditions. NVIDIA says the model supports synchronized video from up to 11 camera views, including fisheye cameras, and RGB–LiDAR generation. Optional initial RGB images can help determine the scene’s appearance. LiDAR generation also uses aligned HD map range-map controls. World Scenario Map (WSM) controls guide scene structure and traffic movement, while text prompts guide appearance.
Cosmos3-Nano-Sim-Auto is a causally distilled version of Transfer-Auto for streaming simulation. According to NVIDIA, the model uses current driving controls and previous generation history to produce camera video and LiDAR observations in short chunks. Closed-loop interaction also requires a simulation runtime and a WSM generator to update control data.
Cosmos3-Nano-Sim-Transfer focuses on generating RGB images from simulator controls. Information such as depth, segmentation, edges and blur can be provided as controls; physical state and motion calculations remain in the simulator. The fourth model, Sim-Bimanual, predicts future observations from robot or camera actions. The announcement includes examples of bimanual interaction for YAM and AgiBot robot configurations.
License and setup information
NVIDIA describes the four checkpoints as post-training examples for Cosmos 3 Nano. The shared training approach includes domain and control adaptation, teacher-forced training, consistency distillation and training with the model’s own generated history. The company recommends testing model outputs against the data used for training and evaluation; a successful short test run alone does not show that training is complete.
According to the announcement, the models are subject to the terms of OpenMDW1.1, which permits commercial and non-commercial use. For Sim-Bimanual, setup requirements are listed as Linux, an NVIDIA GPU, Docker and NVIDIA Container Toolkit. The published vLLM-Omni setup was tested on a single GB200; this is not stated as a minimum memory requirement or a mandatory GPU tier.
Impact on architecture and visualization workflows
Cosmos 3 models are presented for driving simulation and robotics, not as tools for architectural modeling or interior visualization. Still, the approach of turning simulator controls into images may interest technical teams exploring visual data and interaction in controlled virtual scenes. The announcement does not describe direct integration with architectural design or archviz software.
For offices and visualization teams, any potential benefit is therefore more likely to be explored in simulation-based research than through immediate integration into a production pipeline. Before deployment, teams should review the requirements of the model they plan to use, its compatibility with their target hardware and the terms of OpenMDW1.1. Adapting it to a specific sensor setup or operating environment may also require post-training.
Sources
1 sourceSource texts are not republished; short quotes are marked, everything else is our own summary and commentary.
The main takeaway from the Cosmos 3 announcement is that, rather than offering a single model, it provides separate options for tasks such as offline driving data generation, streaming simulation and visual prediction of robot movements. This range is relevant to teams working in simulation and AI research, but it is not presented as a new feature for direct use in architectural or archviz production.
For offices and visualization teams in Turkey, evaluation should start with the use case. The setup requirements listed for Sim-Bimanual and the example setup tested on a single GB200 should not be interpreted as minimum hardware requirements for all models. It is also worth considering the license terms and any need to adapt the models with data before testing.
Frequently asked questions
Which models were added to NVIDIA Cosmos 3?
NVIDIA’s announcement lists four models: Transfer-Auto, Sim-Auto, Sim-Transfer and Sim-Bimanual. The company says each has 16 billion trainable parameters.
What license are the Cosmos 3 models available under?
According to NVIDIA, the models are subject to the terms of OpenMDW1.1, which permits commercial and non-commercial use.
What setup requirements were announced for Cosmos 3?
Linux, an NVIDIA GPU, Docker and NVIDIA Container Toolkit are listed in the context of setting up Sim-Bimanual. The published vLLM-Omni setup for this model was tested on a single GB200; the announcement does not define this as a minimum hardware requirement.
