whatmodel.sh / video / Cosmos3-Super
Cosmos3-Super
↗ share
+ contribute
Cosmos3-Super is an open-weights video model from NVIDIA, released May 2026. Its published size is 64B parameters. Licence: openmdw-1.1.
Catalogue identity notes
The mid-trained Cosmos 3 Super omnimodel checkpoint (nvidia/Cosmos3-Super, pinned f543c56), catalogued as a video model like cosmos3-nano. The Cosmos 3 report (arXiv 2606.02800v4) lists 'Cosmos3-Super huggingface.co/nvidia/Cosmos3-Super' among its open-weight checkpoints and says the post-trained Cosmos3-Super-Text2Image and Cosmos3-Super-Image2Video were produced from this mid-trained checkpoint; Table 1 stars (∗) only post-trained variants, and the Text2Video 80.0 / Image2Video 82.8 cells for Cosmos3-Super are unstarred, so Tables 12 and 14 rows 'Cosmos3-Super' are this checkpoint, not the separate Cosmos3-Super-Image2Video repo (not proposed: no scored rows). Repo last modified 2026-09-16; weights may have been refreshed after the v4 report. Distinct from the catalogued cosmos3-super-text2image.
Model information sources
Release date ↗ · Weights licence ↗ · Parameter count ↗
Whole Mixture-of-Transformers model (autoregressive reasoner tower + diffusion generator tower) trainable parameters, as stated by NVIDIA ('Cosmos3-Super: 64B'). HF safetensors header total 64,615,003,632.
$ how it compares
The video models ranked either side of Cosmos3-Super, out of 139 with a score, then more from NVIDIA.
| # | model | score | where | ||
|---|---|---|---|---|---|
| 36 |  Runway Gen-4.5 | 58±2 | api $0.12/sec | ||
| 37 |  Wan 2.1 I2V Plus | 58±5 | api price unknown | ||
| 38 | Cosmos3-Super | 58±4 | local | ||
| 39 |  Seedance 1.5 Pro | 58±2 | api $0.052/sec | ||
| 40 |  Kling 2.5 Turbo Pro | 57±2 | api $0.070/sec | ||
| more from NVIDIA | |||||
| 60 | Cosmos3-Nano | 52±3 | local | ||
| 72 | Cosmos-1.0-Diffusion-7B-Video2World | 49±5 | comparison only | ||
| 74 | Cosmos-Predict2.5-2B | 48±2 | comparison only | ||
$ what it's good at
text to video▆▄▃▄▆▅█▃▂▂▂▄ #22 of 68 * · 51% of its score · 1 test
Trails Kling 2.5 Turbo Pro, Seedance 1.5 Pro here, though they rank lower overall.
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Cosmos HUE human evaluation, text-to-video (Cosmos 3 report) lab | 89.3% ✎ | █▄▄··▄▄▄▄▄·▄ beats 7/9 | 51% |
image to video▁▂·▂▅█▅▄▅▆▃▄ #26 of 50 * · 41% of its score · 1 test
| test | result | vs the others in its run | counts |
|---|---|---|---|
| Cosmos HUE human evaluation, image-to-video (Cosmos 3 report) lab | 89.6% ✎ | ▄·█··▄··▄▄▄█ beats 7/8 | 41% |
$ where to run it
Other ways to run it
- Run it as published · needs ~159.7 GB, NVIDIA only · from huggingface.co/nvidia (official)
uvx --from huggingface_hub hf download nvidia/Cosmos3-Super assets/ --local-dir Cosmos3-Super && cd Cosmos3-Super && uv run --python 3.13 --with git+https://github.com/huggingface/diffusers --with accelerate --with av --with cosmos_guardrail --with huggingface_hub --with imageio --with imageio-ffmpeg --with torch --with torchvision --with transformers python -c "import json, torch; from diffusers import Cosmos3OmniPipeline; from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler; from diffusers.utils import export_to_video; p=Cosmos3OmniPipeline.from_pretrained('nvidia/Cosmos3-Super', torch_dtype=torch.bfloat16, device_map='cuda', enable_safety_checker=True); p.scheduler=UniPCMultistepScheduler.from_config(p.scheduler.config, flow_shift=10.0); r=p(prompt=json.dumps(json.load(open('assets/example_t2v_prompt.json'))), negative_prompt=json.dumps(json.load(open('assets/negative_prompt.json'))), num_frames=189, height=720, width=1280, num_inference_steps=35, guidance_scale=6.0, generator=torch.Generator(device='cuda').manual_seed(123)); export_to_video(r.video, 'cosmos3_super_t2v.mp4', fps=24)"
Official weights and documentation: huggingface.co/nvidia ↗
$ how the score was worked out · Cosmos 3: Omnimodal World Mod… 92%
Each result implies a score on its own, after adjusting for how hard its test is and who ran it. The score is their weighted average, with a small pull toward the middle for thinly tested models. Each test's weight is its counts column above; here it's added up by who ran the tests. How it works
| source | results | share of the score |
|---|---|---|
 Cosmos 3: Omnimodal World Models for Physical AI (arXiv:2606.02800v4) · lab-reported | 2 | 92% |
| prior · pull toward the middle | 8% | |
| weighted average | = 58.0 |
$ improve this page
Use a short GitHub form or edit the data through a pull request. Every number needs a source; here's which ones we can use.
data updated · sources and licences · json