Models & labs Brief
NVIDIA adds four open-weight Cosmos 3 Nano models for interactive world simulation and driving scenes
NVIDIA announced on 5 November 2026 four new open-weight checkpoints in its Cosmos 3 family, each with 16 billion trainable parameters, for interactive world simulation and multi-sensor driving-scene generation.
This story was written with the help of artificial intelligence, checked automatically, then reviewed by an editor before publication, under our AI policy.
What NVIDIA released
According to a Hugging Face blog post dated 5 November 2026, NVIDIA has added four open-weight checkpoints to its Cosmos 3 family. They are aimed at interactive world simulation and at generating driving scenes from multiple sensors and multiple views.
All four are post-trained variants of Cosmos 3 Nano and have 16 billion trainable parameters. They are published under the OpenMDW1.1 licence, which permits commercial and non-commercial use according to its terms.
Driving models: Transfer-Auto and Sim-Auto
Cosmos3-Nano-Transfer-Auto generates synthetic multi-view driving data with varied weather, lighting and time-of-day conditions. NVIDIA says it produces synchronised video across up to 11 camera views, including fisheye cameras, and supports joint RGB-LiDAR generation.
Cosmos3-Nano-Sim-Auto is the streaming variant, distilled causally from Transfer-Auto. It generates short successive segments of synchronised camera video and LiDAR in a few denoising steps.
Simulation and robotics models
Cosmos3-Nano-Sim-Transfer takes controls from a physics simulator plus a text prompt and outputs photorealistic RGB observations. Physics and ground-truth state stay in the simulator.
Cosmos3-Nano-Sim-Bimanual simulates how an environment evolves in response to the actions of a bimanual robot. It covers the YAM and AgiBot embodiments and can be driven by a policy or by joysticks.
Architecture and deployment notes
NVIDIA states that the Cosmos 3 streaming generation flow uses a Mixture-of-Transformers backbone and a four-stage training process, one stage of which is self-forcing on generated history.
The published vLLM-Omni configuration was tested on a GB200. NVIDIA cautions that "this is not a minimum-memory specification". The post, as summarised here, does not give minimum hardware requirements, and the performance descriptions above are NVIDIA's own claims rather than independent results.
| Model | Stated purpose |
|---|---|
| Cosmos3-Nano-Transfer-Auto | Multi-view synthetic driving data, up to 11 cameras, joint RGB-LiDAR |
| Cosmos3-Nano-Sim-Auto | Streaming, causally distilled variant generating camera and LiDAR segments |
| Cosmos3-Nano-Sim-Transfer | Simulator controls and text prompt to photorealistic RGB |
| Cosmos3-Nano-Sim-Bimanual | Environment simulation for bimanual robots (YAM, AgiBot) |
FAQ
What licence covers the new Cosmos 3 models?
They are released under OpenMDW1.1, which allows commercial and non-commercial use according to its terms.
How large are the models?
Each of the four is a post-trained Cosmos 3 Nano variant with 16 billion trainable parameters.
Which hardware was used to test the vLLM-Omni setup?
A GB200. NVIDIA says this is not a minimum-memory specification.
Sources
- New NVIDIA Cosmos 3 Open Models for Interactive World Simulation, Hugging Face, (primary source)