Nvidia posts four open Cosmos 3 world-simulation checkpoints
The four 16B-parameter Cosmos 3 Nano checkpoints on Hugging Face target world simulation, such as driving and robotics, rather than creative video.
By KreatorFlow News
1 min2 sources

Nvidia published four open-weight Cosmos 3 Nano checkpoints on Hugging Face on November 5, 2026. They are built for world simulation rather than creative video.
What happened
Each checkpoint has 16 billion parameters. The four cover:
- multi-camera driving video
- streamed driving simulation
- depth or outline to video
- bimanual (two-handed) robot action to video
Reports list the OpenMDW-1.1 licence for these checkpoints. Licence details for other Cosmos 3 releases differ, so the terms should be checked for each model.
Running the checkpoints needs high-end hardware: about 96 GB of VRAM, according to one roundup.
Why it matters for creators
World models share the same video-generation foundations as creative models, and they hint at future video that is more controllable and more aware of physics.
These checkpoints are not creative tools. They are aimed at training and testing systems such as self-driving cars and robots, and the hardware they need puts them out of reach of most desktop setups. The depth- and outline-to-video checkpoint is the closest to creative workflows, since it turns structural guides into video.
Sources
What this story is based on. Links open the original pages.
- NVIDIA on Hugging FaceCosmos interactive world simulation
- Let's Data ScienceNVIDIA Releases Four Open Cosmos 3 Checkpoints for World Simulation


