whatmodel.sh / video / Cosmos3-Nano

Cosmos3-Nano

+ contribute
Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

Cosmos3-Nano is an open-weights video model from NVIDIA, released May 2026. Its published size is 16B parameters. Licence: openmdw-1.1.

Model information sources

Release date ↗ · Weights licence ↗ · Parameter count ↗

Whole Mixture-of-Transformers model (autoregressive tower + diffusion tower) trainable parameters, as stated by NVIDIA.

51.8±2.6#48 of 93 video modelsfrom 6 benchmark results across 5 sources

$ how it compares

The video models ranked either side of Cosmos3-Nano, out of 139 with a score, then more from NVIDIA.

#modelscorewhereRAMreleased
58HY-OmniWeaving Tencent52±4comparison only2026-03
59HunyuanVideo 1.5 Tencent52±2api $0.075/sec +local~20GB2025-11
60Cosmos3-Nano NVIDIA52±3local~42GB2026-05
61PixVerse 5.5 PixVerse52±10api $0.060/sec2025-12
62Kling 1.6 (Pro mode) Kuaishou (Kling AI)52±4comparison only2024-12
more from NVIDIA
38Cosmos3-Super NVIDIA58±4local~160GB2026-05
72Cosmos-1.0-Diffusion-7B-Video2World NVIDIA49±5comparison only2025-01
74Cosmos-Predict2.5-2B NVIDIA48±2comparison only2025-10

see all video models · find the ones that fit your machine

$ what it's good at

text to video▆▄▃▄▆▅█▃▂▂▂▄ #29 of 68 * · 32% of its score · 2 tests

Beats Kling 2.1 Master, HunyuanVideo 1.5 here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
Cosmos HUE human evaluation, text-to-video (Cosmos 3 report) lab87.6% ✎█▄▄··▄▄▄▄▄·▄ beats 5/9easy · avg 85%▇▇▇▇▇▇▇▇▇ 23%typical
Sci-VBench testmini (expert human rating)44.6% ✎▅·▃█▃·▅▃▅▃▃▃ beats 4/14hard · avg 44%▇▇▇▇ 9.3%blunt · at its level
image to video▁▂·▂▅█▅▄▅▆▃▄ #27 of 50 * · 22% of its score · 2 tests

Beats Seedance 1.5 Pro, Kling 2.5 Turbo Pro, Wan 2.2 I2V A14B here, though they rank higher overall.

testresultvs the others in its rundifficultycountswhy
Cosmos HUE human evaluation, image-to-video (Cosmos 3 report) lab88.6% ✎▄·█··▄··▄▄▄█ beats 6/8easy · avg 84%▇▇▇▇▇▇▇ 18%typical
Physics-IQ Verified (I2V, best-practice prompts)29.1% ✎██···█··█·██ beats 2/5hard · avg 25%▇▇ 4.6%blunt · at its level
other tests2 not in a task · 43% of its score
testresultvs the others in its rundifficultycountswhy
Rapidata relative camera movement (I2V, alignment Elo)717.5 ✎█···▄▄██·█▄█ beats 1/12avg 661▇▇▇▇▇▇▇▇▇▇▇ 28%sharp · noisy
Rapidata image-to-video realism (overall Elo)717.1 ✎▃······▃▂▃▃█ beats 0/12avg 871▇▇▇▇▇▇ 15%sharp · noisy

$ where to run it

Other ways to run it

  1. Run it as published · needs ~42.4 GB, NVIDIA only · from huggingface.co/nvidia (official)
    uvx --from huggingface_hub hf download nvidia/Cosmos3-Nano assets/ --local-dir Cosmos3-Nano && cd Cosmos3-Nano && 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-Nano', 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_nano_t2v.mp4', fps=24)"

Official weights and documentation: huggingface.co/nvidia ↗

$ how the score was worked out · Cosmos 3: Omnimodal World Mod… 40%, Rapidata/relative-camera-move… 28%, Rapidata/image-to-video-reali… 15% and 2 more

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

sourceresultsshare of the scorelicence
Cosmos 3: Omnimodal World Models for Physical AI (arXiv:2606.02800v4) · lab-reported2▇▇▇▇▇▇▇▇▇▇▇▇ 40%CC-BY-4.0
Rapidata/relative-camera-movement-human-preference-900k published dataset-card rankings1▇▇▇▇▇▇▇▇ 28%CC-BY-4.0
Rapidata/image-to-video-realism published dataset-card rankings1▇▇▇▇ 15%CC-BY-4.0
Sci-VBench: Evaluating Knowledge- and Reasoning-Intensive Video Generation in Science Domains1▇▇▇ 9%CC-BY-4.0
Physics-IQ Verified1▇ 5%CC-BY-4.0
prior · pull toward the middle▇ 3%
weighted average= 51.8

$ improve this page

Update model informationAdd a place to run itAdd a test resultPropose a benchmarkCorrect a benchmark scoreEdit data on GitHub

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