Skip to main content
Back

NVIDIA's Cosmos3 Tops Open-Source Video Leaderboard With 64B World Model

NVIDIA's 64B Cosmos3-Super-Image2Video lands as an open-weights model that tops the Artificial Analysis image-to-video leaderboard among open systems.

Pro

  • NVIDIA released Cosmos3-Super-Image2Video, a 64B open-weights image-to-video model under OpenMDW 1.1 (commercial use allowed).
  • Tops the Artificial Analysis open-source image-to-video leaderboard; competitive with closed systems on the all-models chart.
  • Built on a Mixture-of-Transformers with a 32B autoregressive reasoner and a 32B diffusion generator sharing joint attention.
  • 50-step generation runs in about 55 seconds on 8xH200, or roughly 170 seconds on a single GB200.
  • Default output is 189 frames at 24 fps; supports 256p/480p/720p across five aspect ratios.
  • Part of the Cosmos 3 family aimed at Physical AI: robotics, AV, and smart-space synthetic data generation.

NVIDIA quietly dropped one of the more interesting open-weights releases of the year: Cosmos3-Super-Image2Video, a 64B-parameter image-to-video generator that ships under a commercial-friendly license and currently sits at the top of the open-source image-to-video leaderboard on Artificial Analysis.

It is part of the broader Cosmos 3 launch, which unifies language, images, video, audio, and actions within a single architecture, enabling developers to build agents that can understand, reason about, simulate, and act in the physical world. The Image2Video variant is the piece most people will reach for first, because it turns a single still frame plus a text prompt into a temporally coherent clip up to several seconds long.

What actually shipped

The model is live on Hugging Face with weights, inference scripts, and example assets. It is released under the OpenMDW 1.1 license and the card says it is ready for commercial and non-commercial use. NVIDIA lists the release date as May 31, 2026 on Hugging Face and GitHub. The weights themselves are free; you pay for the GPUs to run them.

Some concrete specs worth knowing:

  • 65B parameters total, BF16 precision, served as safetensors
  • Resolutions: 256p, 480p, 720p across 16:9, 4:3, 1:1, 3:4, 9:16
  • Default output is 189 frames at 24 fps (roughly 7.8 seconds), with a max of 400 frames
  • Supported hardware: NVIDIA Ampere, Hopper, and Blackwell
  • Two serving paths: vllm-omni with an OpenAI-compatible endpoint, or Hugging Face Diffusers via the new Cosmos3OmniPipeline

On an 8xH200 box with the recommended config, a 50-step generation takes roughly 55 seconds. On a single GB200, the Diffusers path produces a clip in about 170 seconds. A 2xH200 fallback exists but stretches to about 3 minutes per video.

How to call it

The Diffusers integration is the path of least resistance for anyone who has run a Stable Diffusion or Wan pipeline before. The minimal script looks like this:

import torch
from diffusers import Cosmos3OmniPipeline, UniPCMultistepScheduler
from diffusers.utils import export_to_video, load_image
pipe = Cosmos3OmniPipeline.from_pretrained(
    "nvidia/Cosmos3-Super-Image2Video",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)
pipe.scheduler = UniPCMultistepScheduler.from_config(
    pipe.scheduler.config, flow_shift=5.0
)
image = load_image("first_frame.png")
result = pipe(prompt=prompt, image=image, num_frames=189,
              height=480, width=832, num_inference_steps=50,
              guidance_scale=6.0)
export_to_video(result.video, "out.mp4", fps=24)