Cosmos 3
NVIDIA's open world models for robots and driving, in 4B, 16B and 64B sizes, that read and write text, images, video and actions.
In short. Cosmos 3 is a family of AI models from NVIDIA that understand and simulate the physical world. You can give it text, pictures, video or robot movements, and it can answer in words, draw an image, generate a short video, or predict what a robot should do next. The two larger models also handle sound. It is built for robotics and self-driving research, and the license is permissive.
Summary
NVIDIA calls Cosmos 3 a family of omnimodal world models for Physical AI, meaning robots, autonomous vehicles and smart spaces. One model takes text, images, video and robot action trajectories as input, plus audio in the larger two, and writes text, images, video and actions, plus audio. Cosmos3-Nano, with 16 billion parameters, and Cosmos3-Super, with 64 billion, were published on Hugging Face on 31 May 2026, and the 4B Cosmos3-Edge followed on 20 July 2026 without audio. The design is a Mixture-of-Transformers with two towers in one network: an autoregressive transformer that writes text and reasoning, and a diffusion transformer that generates images, video, audio and actions by iterative denoising.
The generator makes video of 5 to 400 frames at 256p, 480p or 720p, 189 frames by default, with 48 kHz stereo audio muxed into the MP4, and it makes images and action sequences for supported embodiments such as camera motion, cars, egocentric views, Franka arms, AgiBot, UR, Google robot, WidowX and UMI. The reasoner reads images and video with up to 256K tokens of context and answers in text with reasoning, 2D and 3D points and bounding boxes. Text prompts are limited to 4,096 tokens, and NVIDIA recommends expanding short prompts into a JSON structure with an LLM first. Its example script calls an LLM endpoint you choose and uses Claude Opus 4.6. NVIDIA also publishes post-trained checkpoints for image-to-video, text-to-image, 4-step generation, simulation, transfer and robot policies, which we do not list one by one.
Only BF16 is officially tested, and the cards say FP8, FP4 and FP16 are not officially supported. The repositories are about 9.2 GB for Edge, 35 GB for Nano and 133 GB for Super, and the models run on Linux with NVIDIA Ampere, Hopper or Blackwell GPUs through vLLM-Omni, Diffusers, SGLang or PyTorch. NVIDIA's recommended setups are one H200 for Nano and eight H200, H100 or A100 GPUs for Super, where a 50-step video takes about 55 seconds on H200s, or about 3 minutes on two. For Edge, NVIDIA's tables show a 189-frame 480p image-to-video clip in about 28 seconds on an H100, 166 seconds on a DGX Spark and 138 seconds on a Jetson AGX Thor with vLLM-Omni.
The weights use the permissive OpenMDW-1.1 license, and the cards say the models are ready for commercial and non-commercial use. The cards list 1.3 billion training samples from 393 datasets, including 340 million YouTube video samples, 100 million Coyo-700M images, about 29 million synthetic images made with HiDream-I1 and Qwen-Image-2512 and 1.1 billion synthetic captions written with Qwen3-VL, and say that restricted content was filtered out. Benchmark results appear in charts in the cards and the technical report. NVIDIA says Edge delivers the highest generation throughput with competitive quality on PAIBench, RBench and PhysicsIQ, which is the developer's claim. In the 30 days to 11 November 2026, Hugging Face counted about 1.31 million downloads of Nano, 57,000 of Edge and 43,000 of Super.
Good for
- Robotics and autonomous-driving research that needs a world model to predict video, actions or outcomes
- Generating short physical-scene videos with synchronized audio on your own NVIDIA hardware
- Visual reasoning over video, with points and boxes in the answer
- Starting a policy or simulator fine-tune from a permissively licensed base, using the 4B Edge on a single GPU
Watch out for
- It is built for Physical AI and tops out at 720p and 400 frames, so it is not a general creative video tool
- Only BF16 is officially tested, and Super needs about 133 GB of weights and eight data-center GPUs in NVIDIA's recommended setup
- It runs on Linux with NVIDIA GPUs only
- The best quality needs an LLM to expand prompts into a JSON structure first
- The card's training list includes large scraped sets such as YouTube video, and OpenMDW does not clear third-party rights for you
- The benchmark claims are NVIDIA's own
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Cosmos 3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Cosmos3-Edge BF16 | 3.86B | 9.2 GB | 11 GB | Laptop or phone |
| Cosmos3-Nano BF16 | 15.8B | 35 GB | 40 GB | Workstation |
| Cosmos3-Super BF16 | 64.6B | 133 GB | 153 GB | Big-memory desktop |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs.
See also
- Cosmos 3 vs Wan 2.2: which open video model to pick?
- Best open-source image and video generation models
- Cosmos 3: NVIDIA's open world models for robots and driving
Frequently asked questions
- What is Cosmos 3?
- Cosmos 3 is a family of AI models from NVIDIA that understand and simulate the physical world. You can give it text, pictures, video or robot movements, and it can answer in words, draw an image, generate a short video, or predict what a robot should do next. The two larger models also handle sound. It is built for robotics and self-driving research, and the license is permissive.
- Can I use Cosmos 3 commercially?
- Yes. Cosmos 3 is released under OpenMDW-1.1, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is Cosmos 3 open source?
- Cosmos 3 is open-weight, not open source in the strict sense. You can download the model files, but the OpenMDW-1.1 adds conditions that the Open Source Initiative definition does not allow.
- Can I run Cosmos 3 on my own computer?
- Yes. The smallest variant, Cosmos3-Edge, needs about 11 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Cosmos 3 released?
- Cosmos 3 was released by NVIDIA on May 31, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- 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
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| 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 | |
| Pika H3 Flex Pika | 33.1B | – | MiniMax H3 Community (Conditional) | 169 GB Big-memory desktop |