Cosmos 3: Omnimodal World Models for Physical AI

Model Collection | Code | White Paper | Website

NVIDIA Cosmos™ is a world foundation model platform designed to accelerate the development of Physical AI by enabling machines to understand, simulate, and interact with the physical world across robotics, autonomous driving, and smart space environments, including industrial and factory-scale applications.

Model Overview: Cosmos3-Super-Text2Image-4Step

The highlighted checkpoint, Cosmos3-Super-Text2Image-4Step, is a step-distilled variant of Cosmos3-Super-Text2Image.

Description

Cosmos3-Super-Text2Image-4Step is a 4-step distilled version of the base Cosmos3-Super-Text2Image model. Given a text prompt, it generates a high-fidelity image.

• What's different from the base model: Distilled into a fixed 4-step sampling schedule without the need for classifier-free guidance. At the recommended 50-step + CFG setting for the Text2Image model, this replaces 50 × 2 = 100 model evaluations with 4, leading up to 25× fewer diffusion-model evaluations per image (see PBR for latency estimates).
• Quality: At just 4 steps, ranks #2 among open-source models on the Artificial Analysis Text-to-Image leaderboard, within one place of the full-step Cosmos3-Super-Text2Image (as of 2026/07/17; see Benchmarks section).
• Input / Output: Text prompt in → PNG image out. Supports various aspect ratios and resolutions.
• Example usage and output: See Usage section.
• Hardware: The 64B model needs a multi-GPU H100/H200 node (4–8 GPUs) or a single B200; it does not fit on a single smaller GPU. See Usage. Full model details below.

Cosmos3-Super-Text2Image-4Step was developed by NVIDIA as a part of Cosmos3.

Cosmos3 is a collection of Omnimodal world models capable of generating dynamic, high-quality video, image, audio, and action commands from combinations of text, image, video, and action trajectory inputs. It serves as a foundational building block for a broad range of Physical AI applications and research spanning world understanding, world generation, simulation, and embodied policy learning.

This model is ready for commercial and non-commercial use.

Model Developer: NVIDIA

Model Versions

Released on: 07/20/2026

• Cosmos3-Super-Text2Image-4Step:
  • Given text input, generate high-fidelity images that are consistent with the provided description. Distilled from Cosmos3-Super-Text2Image using Improved Distribution Matching Distillation (DMD2), enabling high-quality generation in 4 steps.

License

This model is released under the OpenMDW1.1

Deployment Geography

Global

Use Case

Physical AI: Encompassing robotics, autonomous vehicles (AV), and smart space environments, including industrial and factory-scale applications.

Release Date

Hugging Face 07/20/2026 via https://huggingface.co/collections/nvidia/cosmos3. GitHub 07/20/2026 via https://github.com/nvidia/cosmos.

Model Architecture

Architecture Type: Transformer

Network Architecture: Mixture-of-Transformers (MoT)

Cosmos3 is an Omni-modal foundation model 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, such as 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.

This model was developed based on: Cosmos Framework

Number of trainable model parameters:

Released on: 07/20/2026

• Cosmos3-Super-Text2Image-4Step: 64B

Input/Output Specifications

• Generator Input
  • Input Type(s): Text, Image, Video (with audio or without audio), Action Trajectory
  • Input Format(s):
    • Text: String
    • Image: jpg, png, jpeg, webp
    • Video (with or without audio): mp4
    • Action: json (1D list)
  • Input Parameters:
    • Text: One-dimensional (1D)
    • Image: Two-dimensional (2D)
    • Video: Three-dimensional (3D)
    • Audio: One-dimensional (1D)
    • Action trajectory: One-dimensional (1D)
  • Other Properties Related to Input:
    • For video inputs, we accept various resolutions, including 720p, 480p, and 256p.
    • When using input video with audio muxed into the video MP4 file, the audio should have 2 channels (stereo) and a 48 kHz sample rate.
    • Image and video inputs are RGB color (8 bits per channel, sRGB color space); grayscale inputs are not supported.
    • Action input is a per-frame sequence of robot/agent state or control values (e.g., joint positions, gripper state, camera pose). The full input is a 2D array shaped (T, D), where T is the number of frames and D is the embodiment-specific dimensionality listed below.
    • Input action is only supported for compatible embodiments, including general camera motion (9D), autonomous vehicle (9D), egocentric motion (57D), single Franka Panda arm with RobotiQ gripper (10D), dual Franka Panda arm with RobotiQ gripper (20D), Agibot (29D), UR (10D), Google robot (10D), WidowX 250 (10D), UMI (9D).
  • Input Size and Length limits:
    • Text: 4096 tokens
    • Image: 256p, 480p, and 720p resolution at one of these aspect ratios (16:9, 4:3, 1:1, 3:4, 9:16)
    • Video: 256p, 480p, and 720p resolution at one of these aspect ratios (16:9, 4:3, 1:1, 3:4, 9:16). Max number of frames = 5.
    • Audio: Max 0.5 second
    • Action: 16 – 400 video frames
• Generator Output
  • Output Type(s): Image, video, audio, action, text
  • Output Format(s):
    • Image: JPG
    • Video: MP4
    • Audio: Advanced Audio Coding (AAC) stream (muxed within the MP4)
    • Action: 1D list (.json)
    • Text: string
  • Output Parameters:
    • Image: Two-dimensional (2D)
    • Video: Three-dimensional (3D)
    • Audio: One-dimensional (1D)
    • Action: One-dimensional (1D)
    • Text: One-dimensional (1D)
  • Other Properties Related to Output:
    • The generated video is an MP4 file, with the resolution, frame rate, and duration specified in the input. The generated audio is encoded in AAC format, muxed into the video MP4 file with 2 channels (stereo) and a 48 kHz sample rate.
    • Video generation supports durations from 5 to 400 frames, with 189 frames as the default generation duration.
    • The generated action is only supported for compatible embodiments, including general camera motion (9D), autonomous vehicle (9D), egocentric motion (57D), single Franka Panda arm with RobotiQ gripper (10D), dual Franka Panda arm with RobotiQ gripper (20D), Agibot (29D), UR (10D), Google robot (10D), WidowX 250 (10D), UMI (9D).
    • Audio: 48 kHz stereo AAC stream muxed into video mp4
    • Video: mp4 at the FPS specified in input
    • Image: JPEG
• Reasoner Input
  • Input Type(s): Text, Text+Image, Text+Video
  • Input Format(s):
    • Text: String
    • Image: jpg, png, jpeg, webp
    • Video: mp4
  • Input Parameters:
    • Text: One-dimensional (1D)
    • Image: Two-dimensional (2D)
    • Video: Three-dimensional (3D)
  • Other Properties Related to Input:
    • Video inputs are recommended at a frame rate of 4 fps.
    • Long-context inputs supported up to 256K tokens.
  • Input Size and Length limits:
    • Text: Up to 256K tokens (context window).
    • Image: Standard input image formats; passed as file or URL.
    • Video: mp4 at the recommended 4 fps.
• Reasoner Output
  • Output Type(s): Text
  • Output Format(s):
    • Text: string
  • Output Parameters:
    • Text: One-dimensional (1D)
  • Other Properties Related to Output:
    • Default max_tokens=4096+ is recommended for reasoning outputs; longer outputs may be requested.
    • Reasoning outputs may include structured chain-of-thought, 2D/3D point localization, and bounding-box coordinates for vision-based tasks.

The video content visualizes the input text description as a short animated scene, capturing key elements within the specified time constraints.

Our AI models are designed and/or optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g., GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

Software Integration

Runtime Engine(s):

• PyTorch
• vLLM-Omni
• Hugging Face Diffusers

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere
• NVIDIA Blackwell
• NVIDIA Hopper

Operating System(s):

• Linux (We have not tested on other operating systems.)

Note: Only BF16 precision is tested. Other precisions like FP4, FP8, and FP16 are not officially supported.

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

Training, Testing, and Evaluation Datasets

Dataset Overview

• Total Size: 1.3B data points
• Total Number of Datasets: 393 dataset entries
• Dataset partition: Training [100%], Testing [N/A — evaluation benchmarks used separately], Validation [N/A — evaluation benchmarks used separately]
• Time period for training data collection: 2024–2026
• Time period for testing data collection: N/A (standard public benchmarks)
• Time period for validation data collection: N/A (standard public benchmarks)

Raw data from internal and external sources is transformed into training-ready data through multiple stages of curation, filtering, and quality review. Data acquisition spans diverse multimodal sources — robotics, autonomous driving, industrial environments, indoor and outdoor scenes, varied lighting and weather conditions, camera viewpoints, object categories, and human activities — to broaden coverage across Physical AI operating environments. Automated filtering pipelines remove corrupted, duplicate, low-quality, and restricted content. Metadata analysis, heuristic rules, and model-assisted classifiers are applied during preprocessing to flag anomalous distributions and low-diversity subsets. Human review supplements automated filtering for selected datasets, benchmark construction, and targeted quality analysis. Datasets are balanced across modalities and task categories — visual reasoning, text-to-image, text-to-video, image-to-video, audio generation, video transfer, action-conditioned generation, and action command generation — to reduce overrepresentation of narrow domains. Synthetic and simulation-based augmentation supplements coverage of rare physical interactions and edge-case scenarios. Deduplication and provenance tracking are applied across the corpus. The resulting processed data is converted into model-ready tokenized or encoded representations through modality-specific preprocessors before training begins.

Training datasets passed through multiple layers of automated and manual safeguards designed to reduce the presence of harmful or policy-violating content across categories including weapons and weapons-related instructional content, criminal planning, child sexual abuse material (CSAM), non-consensual intimate imagery (NCII), sexual content involving minors, harassment, hate speech, profanity, threats and incitement to violence, self-harm or suicide-related content, and graphic violence. Data sources are reviewed for licensing compatibility, provenance, and alignment with internal data governance and safety policies before admission into training corpora. Automated filtering pipelines combine multiple detection strategies: hash-matching against known CSAM and NCII reference databases; classifier-based moderation models trained for explicit sexual content, hate speech, violence, weapons imagery, and other restricted categories; keyword and regex-based screening for criminal-planning, threats, and self-harm phrases in text data; metadata and provenance heuristics for source-level risk signals; and embedding-based anomaly detection to surface samples that fall outside expected distributions. Human review and targeted audits supplement automated filtering for selected datasets, benchmark construction, and safety-sensitive evaluation. For multimodal Physical AI data (robotics, autonomous driving, industrial scenes), additional filtering targets invalid action trajectories, physically implausible interactions, and unsafe control sequences. Synthetic and simulation-generated data are evaluated through internal validation before inclusion. Benchmark evaluations and red-team testing are applied post-training to surface remaining safety gaps across world generation, reasoning, audio, and action tasks. No large-scale data-filtering process can guarantee complete removal of all harmful content; residual risks may remain, particularly in rare edge cases or open-world deployment settings. Ongoing monitoring and dataset review continue post-release.

• For more information about the datasets used to train this model, please see the Public Summary of Training Content.

Data Modality and Training Data Size

Data Collection Method by dataset

• Hybrid: Automatic/Sensors, Synthetic, Automated

Labeling Method by dataset

• Hybrid: Human, Automated

Properties: The training, testing, and evaluation datasets consist of diverse multimodal video, image, audio, action, synthetic, and sensor-conditioned data sourced from NVIDIA-owned data and publicly available, commercially permissive datasets. These datasets are curated to exclude known restricted content and to support building an Omni model that learns to generate and reason about dynamic physical environments across world reasoning and generation tasks.

Public Datasets

Private Datasets

Synthetic Datasets

Evaluation Datasets

Data Collection Method by dataset

• Hybrid: Automatic/Sensors, Synthetic, Automated

Labeling Method by dataset

• Hybrid: Human, Automated

Properties: The training, testing, and evaluation datasets consist of diverse multimodal video, image, audio, action, synthetic, and sensor-conditioned data sourced from NVIDIA-owned data and publicly available, commercially permissive datasets. These datasets are curated to exclude known restricted content and to support building an Omni model that learns to generate and reason about dynamic physical environments across world reasoning and generation tasks.

Benchmarks

Text-to-Image Benchmark Results

Artificial Analysis Leaderboard

Open-Source Models [2026/07/17]

All Models [2026/07/17] (Including Closed-Source)

PBR (Performance Benchmark Reporting)

The following tables report estimated inference latency for Cosmos3-Super-Text2Image-4Step using the available Cosmos3-Super Text-to-Image generator benchmarks. The source benchmarks were collected using BF16 precision and batch size 1.

The original Cosmos3-Super measurements used 50 denoising steps with classifier-free guidance (CFG), requiring two model evaluations per denoising step. Cosmos3-Super-Text2Image-4Step uses a fixed 4-step distilled schedule without CFG (guidance_scale=1.0). The latency estimates below therefore apply a theoretical 25× speedup:

50 denoising steps × 2 CFG evaluations ÷ 4 distilled steps = 25×

These values are theoretical estimates rather than direct measurements of the 4-step checkpoint. Actual end-to-end latency may differ because fixed overheads such as text encoding, VAE processing, GPU communication, and output encoding may not scale linearly with the number of denoising steps.

Text-to-Image Estimated Inference Latency

Latency is reported in seconds. The suffix in each column indicates the number of GPUs used; for example, 480p/4 represents 480p generation using four GPUs. N/A means the configuration is unavailable because the model does not fit in GPU memory. Not tested means no source measurement was reported.

vLLM-Omni

PyTorch

_{Benchmark notes:}

1. Results use BF16 precision and batch size 1.
2. Configurations labeled /1, /4, and /8 use one, four, and eight GPUs, respectively.
3. PyTorch source results report average generation or sampling time and use automatic sharding for multi-GPU execution.
4. vLLM-Omni source results report “Avg. Generation Time” and use HSDP/Ulysses for multi-GPU execution.
5. N/A means the configuration is unavailable because the model does not fit in GPU memory.
6. Not tested means no source measurement was reported.

Usage: Prepare Prompts with Prompt Upsampling

Cosmos3-Super-Text2Image-4Step supports dense natural-language prompts directly. It also supports JSON-format upsampled prompts, which can further improve generation quality. Description and code can be found here.

To create a text-to-image upsampled prompt using Opus-4.7:

git clone https://github.com/NVIDIA/cosmos-framework.git packages/cosmos-framework
pip install -e packages/cosmos-framework

export PROMPT_UPSAMPLER_ENDPOINT_URL="https://api.anthropic.com/v1/"
export PROMPT_UPSAMPLER_MODEL_NAME="claude-opus-4-7"
export PROMPT_UPSAMPLER_API_TOKEN="<your_token>"

python -m cosmos_framework.inference.prompt_upsampling \
    --input assets/original_prompt.txt \
    --output /tmp/upsampled_t2i_opus/ \
    --mode text2image \
    --endpoint-url "${PROMPT_UPSAMPLER_ENDPOINT_URL}" \
    --model "${PROMPT_UPSAMPLER_MODEL_NAME}" \
    --api-token "${PROMPT_UPSAMPLER_API_TOKEN}" \
    --resolution 768 \
    --aspect-ratio "1,1"

The JSON-upsampled version of assets/original_prompt.txt is saved in assets/example_caption.json for convenience, and is used for the image generation examples below.

Usage: Run Inference with PyTorch

For PyTorch inference with Cosmos Framework, see Run with Cosmos Framework.

Usage: Run Inference with vLLM-Omni

Container

docker pull vllm/vllm-omni:cosmos3

Multi-GPU Serving

You can use the release-tested vllm-omni package for deploying an OpenAI-compatible API inference endpoint. The following configuration uses 4 GPUs and shards the 64B model with HSDP. This configuration is suitable for H100/H200-class multi-GPU nodes when sufficient memory is available:

vllm serve nvidia/Cosmos3-Super-Text2Image-4Step \
  --omni \
  --host 0.0.0.0 \
  --port 8000 \
  --cfg-parallel-size 2 \
  --ulysses-degree 2 \
  --tensor-parallel-size 1 \
  --use-hsdp \
  --hsdp-shard-size 4 \
  --init-timeout 1800

For an 8-GPU node, use --ulysses-degree 4 --hsdp-shard-size 8. Setting --enable-layerwise-offload can help with memory usage on GPUs with less available memory; however, please note that for text-to-image generation, this may incur a significant performance penalty.

Single-GPU Serving

On a single GPU with sufficient memory, such as B200, serve the model without distributed parallelism or HSDP flags:

vllm serve nvidia/Cosmos3-Super-Text2Image-4Step \
  --omni \
  --host 0.0.0.0 \
  --port 8000 \
  --init-timeout 1800

Single-GPU serving requires enough device memory for the 64B model. GPUs with less memory, such as H100/H200-class devices, may run out of memory; use the multi-GPU configuration above for those cases.

Cosmos3-Super-Text2Image-4Step uses a fixed 4-step distilled schedule. vLLM-Omni detects the distilled checkpoint from scheduler/scheduler_config.json and reads the schedule from fixed_step_sampler_config.t_list. num_inference_steps and flow_shift are not user-configurable request fields for this model. Use guidance_scale=1.0.

Download Example Prompts and Scripts

The inference scripts (scripts/) and example inputs (assets/) live in this model repo. Download just those folders with the Hugging Face CLI:

pip install -U "huggingface_hub[cli]"
hf download nvidia/Cosmos3-Super-Text2Image-4Step scripts/ assets/ \
    --local-dir Cosmos3-Super-Text2Image-4Step
cd Cosmos3-Super-Text2Image-4Step

Run all commands below from the downloaded repo root.

Example: Text to Image Generation

Generate an image from a JSON-format prompt by calling the vLLM-Omni endpoint. By default, the script uses http://localhost:8000, which matches the serving commands above. If the server runs on another host, pass --endpoint http://<server-host>:8000.

python scripts/gen_image.py \
    --prompt-file assets/example_caption.json \
    --output-path scripts/output.png

Or, as a minimal standalone script:

import base64
import json
import requests

# 1. Read JSON-upsampled prompt.
json_prompt = json.load(open("assets/example_caption.json"))
json_prompt["resolution"] = {"H": 768, "W": 768}
json_prompt["aspect_ratio"] = "1,1"

# 2. Build the request body.
request_body = {
    "prompt": json.dumps(json_prompt),
    "size": "768x768",
    "n": 1,
    "guidance_scale": 1.0,
    "negative_prompt": "",
    "seed": 1143,
    "extra_args": {
      "use_resolution_template": False,
      "guardrails": True,
    },
}

# 3. Send the POST request.
url = "http://localhost:8000/v1/images/generations"
print("Sending request to server...")
response = requests.post(url, json=request_body, headers={"Content-Type": "application/json"})
response.raise_for_status()

# 4. Extract the base64 data and decode it into an image.
response_json = response.json()
b64_data = response_json["data"][0]["b64_json"]
image_bytes = base64.b64decode(b64_data)

# 5. Save the final PNG file.
with open("/tmp/cosmos3_t2i_4step.png", "wb") as image_file:
    image_file.write(image_bytes)
print("Saved image to /tmp/cosmos3_t2i_4step.png")

Example output generated from assets/example_caption.json:

Usage: Run Inference with Diffusers

Cosmos3-Super-Text2Image-4Step is supported by the Hugging Face Diffusers modular pipeline. The distilled checkpoint is not compatible with Cosmos3OmniPipeline or Cosmos3OmniModularPipeline; it must be loaded with Cosmos3DistilledModularPipeline.

Install

Install Diffusers from the main branch until a release containing Cosmos3DistilledModularPipeline is available:

uv venv --python 3.13 --seed --managed-python
source .venv/bin/activate
uv pip install \
  "diffusers @ git+https://github.com/huggingface/diffusers.git" \
  accelerate \
  av \
  cosmos_guardrail \
  huggingface_hub \
  imageio \
  imageio-ffmpeg \
  torch \
  torchvision \
  transformers

The safety checker requires access to the gated nvidia/Cosmos-1.0-Guardrail repository.

Example: Text to Image Generation

import json

import torch
from diffusers import Cosmos3DistilledModularPipeline

json_prompt = json.load(open("assets/example_caption.json"))
json_prompt["resolution"] = {"H": 768, "W": 768}
json_prompt["aspect_ratio"] = "1,1"

pipe = Cosmos3DistilledModularPipeline.from_pretrained("nvidia/Cosmos3-Super-Text2Image-4Step")
pipe.load_components(torch_dtype=torch.bfloat16)
pipe.enable_safety_checker()
pipe.to("cuda")

images = pipe(
    prompt=json.dumps(json_prompt),
    num_frames=1,
    height=768,
    width=768,
    add_resolution_template=False,
    generator=torch.Generator(device="cuda").manual_seed(1143),
    output="videos",
)
images[0].save("cosmos3_t2i_4step_diffusers.png")
print("Saved image to cosmos3_t2i_4step_diffusers.png")

Single-frame generation uses num_frames=1, and the frame comes back through the pipeline's videos output.

The 4-step schedule and the guidance scale come from the checkpoint: num_inference_steps is fixed to the length of the distilled sigma schedule and guidance_scale is forced to 1.0, so passing any other value raises an error. negative_prompt is ignored, and unlike the base Cosmos3 models there is no scheduler to swap in and no flow_shift to set.

Example output generated from assets/example_caption.json:

Limitations

Cosmos3 may produce imperfect outputs in challenging scenarios. Generation artifacts include temporal inconsistency, unstable camera or object motion, imprecise physical interactions, inaccurate audio-video synchronization, and action-state drift — especially in long-horizon or high-resolution outputs. Reasoning may also be incorrect: object states, causal relationships, spatial geometry, temporal ordering, agent intent, and future outcomes can be misinferred, and complex or long-context inputs may yield hallucinated entities, inconsistent interpretations, or implausible predictions. Because the model lacks an explicit physics simulator, 3D geometry, 4D space-time evolution, object permanence, contact dynamics, and physical laws are only approximated — producing artifacts such as disappearing or morphing objects, unrealistic collisions, and physically implausible motions. Quality further degrades in out-of-distribution environments, safety-critical edge cases, and domains underrepresented in training.

Cosmos3 outputs should not be treated as physically accurate simulation, reliable ground-truth reasoning, or safety-certified decision making. Applications involving robotics control, autonomous systems, scientific simulation, or safety-critical planning require additional validation, external constraints, system-level safety analysis, and domain-specific guardrails before deployment.

Inference

Acceleration Engine: PyTorch, vLLM-Omni, Hugging Face Diffusers

Test Hardware: RTX PRO 6000, H20, H100, H200, B200

Ethical Considerations

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please make sure you have proper rights and permissions for all input image and video content; if image or video includes people, personal health information, or intellectual property, the image or video generated will not blur or maintain proportions of image subjects included.

Users are responsible for model inputs and outputs. Users are responsible for ensuring safe integration of this model, including implementing guardrails as well as other safety mechanisms, prior to deployment.

For more detailed information on ethical considerations for this model, please see the Model Card++ Explainability, Bias, Safety & Security, and Privacy subcards. Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.