NVIDIA releases 4 open Cosmos 3 models for driving and robots

NVIDIA added 4 open-weight checkpoints (saved models that you can download and adapt) to its Cosmos 3 models. The models make driving scenes and robot video, which developers use to train and test driving and robot software. Each model has 16 billion trainable parameters (values that a model learns), and the terms of use let a person use them for commercial and non-commercial work.

How it works

These are the 4 models:

• Transfer-Auto: makes driving video from a maximum of 11 camera views, with LiDAR (laser distance sensor).
• Sim-Auto: streams camera video and LiDAR for fast, interactive driving simulation.
• Sim-Transfer: turns simulator controls and a text instruction into photorealistic video.
• Sim-Bimanual: predicts video from the controls of a two-arm robot.

The numbers

NVIDIA gave no test results in this release. NVIDIA tells that more work is necessary. It must measure where the models are good for policy training and tests, and where their limits are. It used one GB200 GPU to test its setup. This is not the minimum GPU that can operate the models.

What is not known

The setup must have an NVIDIA GPU and Docker (container software). To use a new sensor setup, a team must first adapt Transfer-Auto with its data. It then makes Sim-Auto from that model.

Insight: Open models that must have an NVIDIA GPU

The Cosmos 3 weights are open, but the setup must have an NVIDIA GPU and NVIDIA software. In October 2026, NVIDIA also gave Cosmos data blueprints that developers get from cloud companies. This looks like a move to put developers on NVIDIA GPUs with free models. For leaders: tell your team which GPUs an open model must have.