In a guide published on Hugging Face on November 5, NVIDIA details four variants of Cosmos 3 Nano for generating driving data and simulating how an environment looks after a robot acts. Its November 2 technical blog already described the weights as available. [1]
What changes
Transfer-Auto varies weather, lighting and time of day in synthetic driving scenes. Sim-Auto generates synchronized camera and lidar observations in successive chunks. With a simulation runtime, it can respond to updated driving controls. [1]
Context
For robotics, Sim-Transfer generates images from spatial inputs such as depth or segmentation maps, while the simulator maintains the physical state. Sim-Bimanual predicts the next visual observation from the actions of a two-armed robot or from camera movement. [1]
Scope and availability
NVIDIA reports 16 billion trainable parameters per variant and distributes the weights under OpenMDW 1.1. The published Sim-Bimanual setup with vLLM-Omni was tested on one GB200. That test describes a specific configuration and leaves minimum memory requirements for other setups undetermined. [1] [2]
What to check in practice
To compare variants for robotics, use the same action sequence and note where the output stops representing the expected movement. Record the hardware, runtime and memory use as well. The published weights make that test possible. The GB200 example does not establish what other hardware can run it.