NVIDIA's Cosmos 3 Super Cuts Video Generation From 50 Steps to 4
NVIDIA's 4-step distilled Cosmos 3 Super models hit #1 open-weights image-to-video and #3 text-to-image on the Artificial Analysis leaderboard, with up to 25x faster inference.
Takeaways
- New #1 open-weights I2V: Cosmos3-Super-Image2Video-4Step tops the Artificial Analysis leaderboard, beating its own 35-step teacher model.
- 25x faster inference: DMD2 distillation cuts denoising from 50 steps to 4 and removes classifier-free guidance entirely.
- Open weights + commercial use: Both models are on Hugging Face under the OpenMDW-1.1 license, free to use commercially.
- Distillation recipe included: NVIDIA open-sourced the full DMD2 training scripts so teams can distill their own fine-tuned Cosmos3 checkpoints.
- Physical AI focus: Designed for robotics simulation, autonomous driving SDG, and warehouse safety data generation -- not primarily a creative tool.
- Known failure modes: Temporal inconsistency, object morphing, and implausible physics remain common issues in complex or long outputs.
NVIDIA just released distilled versions of its 64B Cosmos 3 Super world foundation model, and the benchmarks are turning heads. Cosmos3-Super-Image2Video-4Step and Cosmos3-Super-Text2Image-4Step are now the top-ranked open-weights models in their respective categories on the Artificial Analysis Arena, beating out the full-step teacher models they were distilled from.
50 steps down to 4
The headline number is the step reduction. The distilled models cut sampling from 50 denoising steps to just 4, and remove the need for classifier-free guidance (CFG) -- which otherwise doubles the work per step -- for up to 25x faster inference than the base model, with little to no loss in generation quality. CFG is a technique that runs the model twice per step (once with the prompt, once without) to steer outputs toward the text description; removing it is a significant efficiency win on its own.
The technique behind this is called DMD2 (Improved Distribution Matching Distillation). DMD formulates diffusion distillation as a distribution-level matching problem, training a student generator to match the teacher distribution rather than explicitly imitating teacher sampling trajectories. Its objective combines a distribution matching term with a regression loss for training stability. DMD2 further improves this by eliminating the costly regression dataset, adopting a two-time-scale update for more accurate score estimation, and incorporating adversarial supervision. In plain terms: instead of forcing the student model to retrace the teacher's exact denoising path, DMD2 just asks it to produce outputs that look statistically indistinguishable from the teacher's -- a much looser and more efficient constraint.
What's under the hood
To understand why these distilled models are competitive, it helps to know what Cosmos 3 Super actually is. Cosmos3 is built on a Mixture-of-Transformers (MoT) architecture consisting of two complementary transformer towers: an autoregressive transformer for discrete token generation and a diffusion transformer for continuous multimodal generation. During inference, text is generated through standard next-token autoregressive decoding, while non-text modalities like images, video, audio, and actions are synthesized through iterative denoising. This unified architecture enables Cosmos3 to model heterogeneous modalities within a single framework while preserving generation mechanisms best suited to each modality.
The base Cosmos3-Super model is 64B parameters total -- a dense 32B model paired with a Qwen3-VL 32B, targeting datacenter Hopper and Blackwell GPUs. The 4-step distilled variants share the same architecture; the distillation only changes how many denoising steps are needed at inference time.
On the generation side, the specs are solid:
- Generation supports 256p, 480p, and 720p resolution tiers.
- Frame counts range from 5 to 300, defaulting to 189 -- about 7.9 seconds of video at 24 FPS.
- Sound is generated as stereo AAC at 48 kHz.
Where it lands on the leaderboard
On the Artificial Analysis Image-to-Video leaderboard, the distilled I2V model ranks #1 among open-weight models, ahead of the full Cosmos3 Super teacher it was distilled from. That last point is worth sitting with: the 4-step student outperforms the 50-step teacher in human preference evaluations. On the text-to-image side, Cosmos3-Super-Text2Image-4Step ranks #25 overall and #3 among open-weights models, behind the full agentic Cosmos3-Super-Text2Image (which uses a prompt-expansion pipeline) and HiDream-O1-Image-Dev-2604.
The full leaderboard context from Artificial Analysis:
| Model | Category | Open Weights Rank | Overall Rank | Elo Score |
|---|---|---|---|---|
| Cosmos3-Super-Image2Video-4Step | Image to Video | #1 | #15 | -- |
| Cosmos3-Super-Text2Image-4Step | Text to Image | #3 | #25 | 1,180 |
| Cosmos3-Super-Text2Image (agentic) | Text to Image | #1 | #10 | 1,219 |
What it's built for
Cosmos 3 is not primarily a creative image tool -- it's a physical AI platform. If you're training a robot to fold laundry, building an autonomous driving simulation, or generating synthetic training data for warehouse safety scenarios, Cosmos 3 is the foundation model designed for exactly these use-cases. The idea is to use video generation as imagination to train physical AI without being constrained by what's been physically captured.
Concretely, the 4-step models are useful for:
- Synthetic data generation (SDG): Rapidly generating training images and video clips for robotics and autonomous driving pipelines, where volume matters more than individual render time.
- Simulation at scale: Generating future observations conditioned on robot actions, inferring the actions behind observed demonstrations, and predicting action sequences from current observations and task prompts.
- Fast iteration during post-training: When fine-tuning Cosmos on domain-specific data, faster inference makes the feedback loop much tighter.
- Edge deployment: The reduced compute per image/video makes running these models on constrained hardware more practical.
Known limitations
NVIDIA is transparent about where the model struggles. Cosmos 3 can produce artifacts in long, high-resolution, or physically complex outputs. Common failure modes include temporal inconsistency, unstable camera or object motion, inaccurate sound-video alignment, imperfect action-state consistency, object morphing, inaccurate 3D structure, and implausible physical dynamics. These are the hard cases for any video generation model, and the 4-step distillation doesn't make them easier.
How to get it
Both models are available now on Hugging Face under the OpenMDW-1.1 license, which permits commercial use. On third-party APIs like fal.ai, text-to-image costs $0.04 per generated image, with optional prompt expansion adding $0.02 per request. Image-to-video is billed at $0.05 per second of generated video, rounded up.
NVIDIA is also open-sourcing the full distillation recipe and training scripts at the Cosmos GitHub repository, so developers can distill their own Cosmos3 checkpoints for faster, application-specific image and video generation. That's a meaningful addition: if you've fine-tuned a Cosmos3-Super checkpoint on your own domain data, you can now run the same DMD2 recipe to produce a fast 4-step version of your custom model.
NVIDIA says it will continue advancing Cosmos 3 for physical AI, with upcoming improvements in interactive world generation, driving scenario simulation, and robotics policies, and is investing in faster inference and more efficient post-training across a broader range of hardware. With the distillation recipe now open, the bottleneck for most teams shifts from model access to domain-specific data quality.