Cosmos 3: NVIDIA's open world models for robots and driving
NVIDIA released Cosmos 3 under OpenMDW-1.1: 4B, 16B and 64B world models that generate video, images and robot actions and reason about video, for robotics and autonomous driving.
NVIDIA published Cosmos 3 on Hugging Face in two steps: the 16-billion-parameter Nano and the 64-billion-parameter Super on 31 May 2026, then the 4-billion-parameter Edge on 20 July. NVIDIA calls it a family of omnimodal world models for Physical AI, meaning robots, autonomous vehicles and smart spaces. All three use the permissive OpenMDW-1.1 license.
What is new
The design is a Mixture-of-Transformers with two towers in one network. An autoregressive transformer writes text and reasoning, and a diffusion transformer generates images, video, audio and robot actions by iterative denoising. One model therefore takes text, images, video and action trajectories as input, plus audio in Nano and Super, and writes the same kinds of output.
As a generator it makes video of 5 to 400 frames at up to 720p, 189 frames by default, with 48 kHz stereo sound in the two larger sizes. It also makes images and action sequences for embodiments such as cars, Franka arms, AgiBot and UMI. As a reasoner it reads images and video, with up to 256K tokens of context, and answers in text with points and bounding boxes. NVIDIA recommends expanding short prompts into a JSON structure with an LLM first.
Running it
The repositories are 9.2 GB for Edge, 35 GB for Nano and 133 GB for Super. Only BF16 is officially tested, and the cards say FP8, FP4 and FP16 are not supported. The models run on Linux with NVIDIA Ampere, Hopper or Blackwell GPUs, through vLLM-Omni, Diffusers, SGLang or PyTorch. NVIDIA recommends one H200 for Nano and eight H200, H100 or A100 GPUs for Super. By our estimate Edge needs about 11 GB of memory, Nano about 40 GB and Super about 153 GB.
For Edge, NVIDIA’s tables show a 189-frame 480p image-to-video clip in about 28 seconds on an H100 and 166 seconds on a DGX Spark. In the 30 days to 11 November, Hugging Face counted about 1.31 million downloads of Nano, 57,000 of Edge and 43,000 of Super.
License
OpenMDW-1.1 is a permissive license for models, and the cards say Cosmos 3 is ready for commercial and non-commercial use. There is no revenue cap or territory limit. The cards list 1.3 billion training samples, including 340 million YouTube video samples, so the license does not clear third-party rights in the data for you.
Why it matters
Open video models in our catalog are mostly aimed at creative clips. Wan 2.2 is the permissive one, and LTX-2.5 adds sound. Cosmos 3 aims at a different job, predicting what a scene or a robot will do next, with a reasoning half in the same network. Its limits are 720p video, NVIDIA-only Linux hardware and benchmark claims that come from NVIDIA. See Cosmos 3 vs Wan 2.2 for the trade-offs, and the image and video list for the whole category.
Sources
- Cosmos 3 project page (NVIDIA Cosmos Lab) research.nvidia.com
- Cosmos 3 technical report (NVIDIA Research) research.nvidia.com
- Cosmos3-Nano model card (Hugging Face) huggingface.co
- Cosmos3-Edge model card (Hugging Face) huggingface.co
- Cosmos3-Super model card (Hugging Face) huggingface.co
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Cosmos 3 NVIDIA | 3.86B – 64.6B | – | OpenMDW-1.1 (Permissive) | 11 GB Laptop or phone | |
| Wan 2.2 Alibaba Wan | 5B – 27B 14B active | – | Apache 2.0 (Permissive) | 24 GB Gaming GPU | |
| LTX-2.5 LTX (Lightricks) | 22B | – | LTX-2.x Community (Conditional) | 16 GB Laptop or phone | |
| MiniMax H3 MiniMax | 33.1B | – | MiniMax H3 Community (Conditional) | 166 GB Big-memory desktop | |
| Qwen-Image-2.1 Alibaba Qwen | 7.1B | – | Qwen Research License (Non-commercial) | 20 GB Gaming GPU |