NVIDIA Cosmos

Cosmos is NVIDIA's family of world foundation models: you give it a text description or a video, and it generates a consistent 3D environment with physics, lighting and weather that an autonomous vehicle or robot can be trained inside. The problem it addresses is that collecting real-world driving or manipulation data for rare situations is slow and expensive, so teams generate the rare situations instead. The models are openly licensed. NVIDIA publishes them under the Linux Foundation's OpenMDW 1.1 licence with weights available to download and post-training scripts on GitHub, and offers a hosted catalogue for trying them without local setup. See similar solutions. That is why this sits on our free tools page — but obtaining them free and running them free are different things, and generating video-scale 3D worlds is a substantial GPU workload we have not costed. Correcting this page rather than quietly rewriting it: we previously said Cosmos generates "infinite" environments and "unlimited" synthetic data, that it reduces real-world collection costs "100x", and that it was pre-trained on "20 million hours" of driving data. We read NVIDIA's Cosmos page in full on 7 September 2026 and none of those figures appear on it. The multiplier and the training-hours number were presented here as facts and we cannot source either, so they are gone Read our guide.