# Cosmos

<p align="center">
  <img src="https://github.com/user-attachments/assets/28f2d612-bbd6-44a3-8795-833d05e9f05f" width="274" alt="NVIDIA Cosmos">
</p>


<p align="center">
  <a href="https://www.nvidia.com/en-us/ai/cosmos/">Website</a> |
  <a href="https://github.com/NVIDIA/cosmos-framework">Framework</a>
</p>

## Table of Contents

- [Introduction](#introduction)
- [Cosmos 3](#cosmos-3)
  - [Key Capabilities](#key-capabilities)
  - [Model Architecture](#model-architecture)
  - [Model Family](#model-family)
  - [Supported Generation Settings](#supported-generation-settings)
  - [Input and Output](#input-and-output)
  - [Use Cases](#use-cases)
    - [Generator](#generator)
    - [Reasoner](#reasoner)
  - [Quickstart](#quickstart)
    - [Generator with Diffusers](#generator-with-diffusers)
    - [Generator with vLLM-Omni](#generator-with-vllm-omni)
    - [Reasoner with Transformers](#reasoner-with-transformers)
    - [Reasoner with vLLM](#reasoner-with-vllm)
    - [Reasoner with NIM](#reasoner-with-nim)
  - [Troubleshooting](#troubleshooting)
    - [Which CUDA version should I use?](#which-cuda-version-should-i-use)
    - [Which base container should I use?](#which-base-container-should-i-use)
    - [`torch.cuda.is_available()` is `False`](#torchcudais_available-is-false-the-nvidia-driver-on-your-system-is-too-old)
    - [Import fails with `libxcb.so.1: cannot open shared object file`](#import-fails-with-libxcbso1-cannot-open-shared-object-file)
    - [`uv` errors on install or `sync`](#uv-errors-on-install-or-sync)
  - [Choosing an Integration](#choosing-an-integration)
  - [Examples](#examples)
  - [Inference Benchmarks](#inference-benchmarks)
  - [Finetune](#finetune)
  - [Limitations](#limitations)
- [Ecosystem](#ecosystem)
- [News](#news)
- [License and Contact](#license-and-contact)

## Introduction

NVIDIA Cosmos is an open platform of world models, datasets, and tools that enables developers to build Physical AI for robots, autonomous vehicles, smart infrastructure, and more.

## Cosmos 3

**Cosmos 3** is our newest model family [[Report]](https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf) [[Website]](https://research.nvidia.com/labs/cosmos-lab/cosmos3/). It is a suite of omnimodal world models designed to jointly process and generate language, images, video, audio, and action sequences within a unified Mixture-of-Transformers architecture. By supporting highly flexible input-output configurations, it seamlessly unifies critical modalities for Physical AI — effectively subsuming vision-language models, video generators, world simulators, and world-action models into a single framework.

Cosmos 3 exposes two runtime surfaces:

| Surface | Inputs | Outputs | Use Cases |
|----------|----------|----------|----------|
| **Reasoner** | Text, vision | Text | World understanding, grounding, physical reasoning, task planning, action forecasting, embodied agent reasoning, and autonomous system decision making |
| **Generator** | Text, vision, sound, action | Vision, sound, action | World generation, world simulation, future prediction, synthetic data generation, policy learning, and robot training |

### Key Capabilities

- **World understanding:** Analyze videos and images for captions, temporal events, next actions, spatial grounding, physical plausibility, and causal outcomes.
- **World generation:** Produce images, videos, synchronized sound, and action-conditioned rollouts from text, image, video, or action inputs.
- **Action modeling:** Predict policy actions, inverse dynamics, and forward dynamics for robotics, camera motion, egocentric motion, and autonomous-driving settings.
- **Research and production paths:** Use Diffusers and Transformers for Python-first development, then vLLM-Omni and vLLM for OpenAI-compatible serving.
- **Post-training recipes:** Adapt vision, action, and reasoner workflows with Cosmos Framework training recipes and task-specific evaluation [Coming Soon].

### Model Architecture

![Cosmos 3 model architecture](cookbooks/cosmos3/cosmos3-model-architecture.png)

Cosmos 3 is an omnimodal world model built on a unified Mixture-of-Transformers (MoT) architecture that combines an autoregressive (AR) transformer for reasoning with a diffusion transformer (DM) for multimodal generation. In Reasoner Mode, language and visual understanding tokens are processed through causal self-attention, enabling next-token prediction for tasks such as perception, planning, and world reasoning. In Generator Mode, noisy image, video, audio, and action tokens are denoised through full attention, allowing the model to jointly generate coherent multimodal outputs. Both modes share the same transformer architecture, multimodal attention layers, and a unified 3D multi-dimensional rotary position embedding (mRoPE) representation that encodes spatial and temporal structure across modalities, enabling consistent reasoning over images, videos, audio streams, and action trajectories.


### Model Family

| Model | Size | Primary Capability |
|---------|---------:|---------|
| **[Cosmos3-Nano](https://huggingface.co/nvidia/Cosmos3-Nano)** | 16B | Compact omnimodal world model for multimodal understanding, world simulation, future prediction, action reasoning, and Physical AI. |
| **[Cosmos3-Super](https://huggingface.co/nvidia/Cosmos3-Super)** | 64B | Frontier-scale omnimodal world model for advanced multimodal understanding, world simulation, future prediction, action reasoning, and Physical AI. |
| **[Cosmos3-Super-Text2Image](https://huggingface.co/nvidia/Cosmos3-Super-Text2Image)** | 64B | High-fidelity text-to-image generation. |
| **[Cosmos3-Super-Image2Video](https://huggingface.co/nvidia/Cosmos3-Super-Image2Video)** | 64B | Temporally coherent image-to-video generation. |
| **[Cosmos3-Nano-Policy-DROID](https://huggingface.co/nvidia/Cosmos3-Nano-Policy-DROID)** | 16B | Vision-language robot policy for DROID manipulation and control. |


### Supported Generation Settings

| Setting           | Supported values                        |
| ------------------| --------------------------------------- |
| Resolution tiers  | 256p, 480p, 720p, default=480p          |
| Aspect ratios     | 16:9, 4:3, 1:1, 3:4, 9:16, default=16:9 |
| Frame rates       | 10, 16, 24, and 30 FPS, default=24      |
| Frame count       | 5 to 300 frames, default=189            |
| Precision         | BF16 tested                             |
| Operating system  | Linux                                   |
| GPU architectures | NVIDIA Ampere, Hopper, and Blackwell    |

### Input and Output

| Spec | Value |
| --- | --- |
| Input types | Text, text + image, text + video, text + image + action |
| Input formats | Text string, JPG/PNG/JPEG/WEBP image, MP4 video, JSON action array |
| Vision conditioning | 720p uses 1280x720, 480p uses 832x480, and 256p uses 320x192. Video conditioning uses 5 frames at the matching resolution. |
| Action conditioning | Supported action dimensions depend on the embodiment, including camera motion (9D), autonomous vehicle (9D), egocentric motion (57D), single-arm robot (10D, DROID/UR/Fractal/Bridge/UMI), dual-arm robot (20D, dual DROID arms), humanoid robot (29D, AgiBot). |
| Output types | Image, video, sound, action state, text |
| Output formats | JPG image, MP4 video, AAC sound stream muxed into MP4, JSON action values, text string |
| Prompt length | Fewer than 300 words is recommended for world-generation prompts |
| Sound output | Stereo AAC at 48 kHz when generated with video |

### Use Cases

#### Generator

Generator examples produce non-text outputs conditioned by text, vision, and action inputs.

| Workflow | Inputs | Outputs | What it demonstrates |
| --- | --- | --- | --- |
| **Text-to-image** | Text | Vision | Robotics laboratory scene generation from a text prompt |
| **Text-to-video** | Text | Vision | Industrial video generation from a dense scene description |
| **Text-to-video with sound** | Text | Vision, sound | Synchronized visual and audio generation |
| **Image-to-video** | Text, image | Vision | Robot manipulation animation from a starting image and prompt |
| **Image-to-video with sound** | Text, image | Vision, sound | Image-conditioned motion with synchronized audio |
| **Video-to-video** | Text, video | Vision | Prompt-guided transformation of a robot manipulation video |
| **Video-to-video with sound** | Text, video, sound | Vision, sound | Prompt-guided transformation of a robot manipulation video |
| **Forward dynamics** | Text, vision, action | Vision | Future-state rollout from action and visual context |
| **Action policy** | Text, vision | Action, vision | Action trajectories and rollout video from context |

Generator prompt upsampling expands short scene descriptions into dense structured prompts. The current examples use these sampling defaults:

| Parameter | Value |
| --- | ---: |
| `max_tokens` | `20000` |
| `temperature` | `0.7` |
| `top_p` | `0.8` |
| `top_k` | `20` |
| `repetition_penalty` | `1.0` |
| `presence_penalty` | `1.5` |
| `seed` | `3407` |

#### Reasoner

Reasoner examples produce text outputs from text and vision inputs. It follows Qwen3-VL-compatible message conventions for image and video inputs.

| Workflow | Inputs | Outputs | What it demonstrates |
| --- | --- | --- | --- |
| **Caption** | Video | Text | Detailed video captioning |
| **Temporal localization** | Video, query | Text or JSON | Event detection, timestamp query, and interval question answering |
| **Embodied reasoning** | Video, question | Text | Next-action prediction for robotics and assisted-task settings |
| **Common-sense reasoning** | Video, question | Text | Physical common-sense judgment with visible context |
| **2D grounding** | Image, prompt | JSON boxes | Bounding-box localization from an image prompt |
| **Describe anything** | Image, marked subjects | JSON or text | Attribute captioning for marked subjects |
| **Action CoT** | Image or video, prompt | Text or JSON | Trajectory prediction and driving-scene chain-of-thought |
| **Physical Plausibility Analysis** | Video, prompt | Label | Physical plausibility classification |
| **Situation Understanding** | Video, question | Text | Situation understanding and likely-next-action prediction |

Reasoner examples use the following sampling settings:

| Parameter | Without reasoning | With reasoning |
| --- | ---: | ---: |
| `top_p` | `0.8` | `0.95` |
| `top_k` | `20` | `20` |
| `repetition_penalty` | `1.0` | `1.0` |
| `presence_penalty` | `1.5` | `0.0` |
| `temperature` | `0.7` | `0.6` |

Use this basic message shape for text + vision requests:

```json
[
  {
    "role": "system",
    "content": [{"type": "text", "text": "You are a helpful assistant."}]
  },
  {
    "role": "user",
    "content": [
      {"type": "video_url", "video_url": "https://example.com/video.mp4"},
      {"type": "text", "text": "List the notable events with approximate timestamps."}
    ]
  }
]
```

For explicit reasoning, append this format instruction to the user prompt:

```text
Answer the question using the following format:

<think>
Your reasoning.
</think>

Write your final answer immediately after the </think> tag.
```

### Quickstart

Before running examples, create a Hugging Face access token and then authenticate locally:

```shell
uvx hf@latest auth login
```

Set `HF_HOME` if you want to use a shared cache or a disk with more space.

#### Generator with Diffusers

<details>
<summary>Expand Diffusers generator setup, example, and modes</summary>

Use HuggingFace Diffusers for Cosmos 3 Generator research, training, and model development. This path loads the full Cosmos 3 checkpoint, including the reasoner path, diffusion generation path, and media tokenizers.

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

`--torch-backend=auto` lets uv detect your NVIDIA driver and install a matching CUDA build of `torch`/`torchvision`. Without it, uv pulls the newest CUDA wheel (currently `cu130`), which fails on pre-CUDA-13 drivers with `The NVIDIA driver on your system is too old` and `torch.cuda.is_available()` returns `False`. Pin an explicit backend instead if you prefer, e.g. `--torch-backend=cu128` for a CUDA 12.8 driver.

A `text-to-video` run takes a while: the first run downloads `Cosmos3-Nano`, and diffusion is compute-heavy, running through every inference step before producing output. Long step times are expected, not a hang.

```python
import torch
from diffusers import Cosmos3OmniPipeline
from diffusers.schedulers.scheduling_unipc_multistep import UniPCMultistepScheduler
from diffusers.utils import export_to_video

pipe = Cosmos3OmniPipeline.from_pretrained(
    "nvidia/Cosmos3-Nano",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
)
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config, flow_shift=10.0)

result = pipe(
    prompt="A mobile robot navigates a warehouse aisle and stops at a shelf.",
    negative_prompt="",
    image=None,
    num_frames=189,
    height=720,
    width=1280,
    fps=24,
    num_inference_steps=35,
    guidance_scale=6.0,
    enable_sound=False,
    add_resolution_template=False,
    add_duration_template=False,
    generator=torch.Generator(device="cuda").manual_seed(1234),
)

export_to_video(result.video, "cosmos3_t2v.mp4", fps=24, macro_block_size=1)
```

Diffusers modes:

| Mode | Use |
| --- | --- |
| `text-to-image` | Single-frame image generation with `num_frames=1`; returns a PIL image |
| `text-to-video` | Video generation; 189 frames is about 7.9 seconds at 24 FPS |
| `image-to-video` | Video generation conditioned on an input image |
| `text-to-video-with-sound` | Video generation with sound for checkpoints that include sound modules |

See the [Cosmos 3 Diffusers documentation](https://huggingface.co/docs/diffusers/main/en/api/pipelines/cosmos3) for runnable examples of each mode.

</details>

#### Generator with vLLM-Omni

<details>
<summary>Expand vLLM-Omni generator setup, endpoints, and request reference</summary>

Use vLLM-Omni for Generator production inference behind an OpenAI-compatible API. This integration loads the full Cosmos 3 checkpoint, including the Qwen3-VL-based reasoner path and the diffusion generation path. For understanding-only tasks that return text, use [Reasoner with vLLM](#reasoner-with-vllm) instead, which loads only the reasoner.

> **Compatibility status:** Cosmos 3 Generator support has landed in [vllm-project/vllm-omni](https://github.com/vllm-project/vllm-omni) `main`: text-to-image, text-to-video, and image-to-video ([#3454](https://github.com/vllm-project/vllm-omni/pull/3454)) and video-with-sound ([#4073](https://github.com/vllm-project/vllm-omni/pull/4073)) are merged; action (policy / forward-dynamics) is in review ([#4102](https://github.com/vllm-project/vllm-omni/pull/4102)) and video-to-video is planned. The `vllm/vllm-omni:cosmos3` Docker image remains the easiest all-in-one build. For current setup and per-modality usage, see the maintained recipes: [Cosmos3-Nano](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-Nano.md) and [Cosmos3-Super](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-Super.md).

Start the server from the Docker image (all modalities). Mount any directory that contains local media or action files you want the server to read.

```shell
docker run --runtime nvidia --gpus all \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -v "$(pwd):/workspace" \
  -p 8000:8000 \
  --ipc=host \
  vllm/vllm-omni:cosmos3 \
  vllm serve nvidia/Cosmos3-Nano \
  --omni \
  --model-class-name Cosmos3OmniDiffusersPipeline \
  --allowed-local-media-path / \
  --port 8000 \
  --init-timeout 1800
```

Cosmos3 checkpoints can exceed the default server init timeout; use
`--init-timeout 1800` on every `vllm serve` command in this section.

vLLM-Omni prints `Application startup complete.` when the API is ready.

For `nvidia/Cosmos3-Super` (the larger 64B model), split weights across GPUs and optionally offload layers to reduce peak memory: `--tensor-parallel-size` splits model weights across multiple GPUs, and `--enable-layerwise-offload` offloads transformer blocks between CPU and GPU with a latency tradeoff and extra CPU RAM use. For example, on four GPUs, add `--tensor-parallel-size 4 --enable-layerwise-offload --init-timeout 1800` to the `vllm serve` command.

Additional parallelism options:

| Option | Use |
| --- | --- |
| `--cfg-parallel-size 2` | Runs the positive and negative CFG branches in parallel on two GPUs. Set CFG strength with the request-level `guidance_scale`; do not use `true_cfg_scale`. |
| `--ulysses-degree 2` | Enables Ulysses sequence parallelism, splitting the sequence dimension across GPUs. |

When combining parallelism options, ensure the server has enough GPUs for the product of the enabled degrees (`tensor_parallel_size` × `cfg_parallel_size` × `ulysses_degree`).

To install vLLM-Omni from `main` instead of using the Docker image (text-to-image, text-to-video, image-to-video, and video-with-sound are merged there; see the [Cosmos3-Nano](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-Nano.md) and [Cosmos3-Super](https://github.com/vllm-project/vllm-omni/blob/main/recipes/cosmos3/Cosmos3-Super.md) recipes for per-modality usage), create a venv and install, choosing the CUDA build that matches your driver:

```shell
uv venv --python 3.13 --seed --managed-python
source .venv/bin/activate
# CUDA 13 driver:
uv pip install --torch-backend=cu130 \
  "vllm-omni @ git+https://github.com/vllm-project/vllm-omni.git@main"
# CUDA 12.8 driver:
# uv pip install --torch-backend=cu128 \
#   "vllm-omni @ git+https://github.com/vllm-project/vllm-omni.git@main"
```

Then run `vllm serve nvidia/Cosmos3-Nano --omni --model-class-name Cosmos3OmniDiffusersPipeline --allowed-local-media-path / --port 8000 --init-timeout 1800` directly, without the `docker run ... vllm/vllm-omni:cosmos3` wrapper.

Vision endpoints:

| Mode | Endpoint | Notes |
| --- | --- | --- |
| Text to image | `POST /v1/images/generations` | Returns a base64-encoded PNG |
| Text to video | `POST /v1/videos/sync` | Blocks and returns the MP4 bytes directly |
| Image to video | `POST /v1/videos/sync` | Upload the conditioning image with `input_reference` |
| Video to video | `POST /v1/videos/sync` | Upload a source video and choose which frames stay as clean conditioning |
| Video with sound | `POST /v1/videos/sync` | Add `generate_sound=true` to produce a soundtrack alongside the video |

Action modes use Cosmos 3 as a world model: they condition on an embodiment (`domain_name`) and exchange video and action sequences. Policy and inverse dynamics return a predicted action chunk, so send those through the asynchronous `POST /v1/videos` job and read the action data from the completed result; forward dynamics returns only video and can use synchronous `POST /v1/videos/sync`.

| Mode | `action_mode` | Input | Output |
| --- | --- | --- | --- |
| Policy | `policy` | Image + instruction | Video + predicted action chunk |
| Inverse dynamics | `inverse_dynamics` | Video + instruction | Video + predicted action chunk |
| Forward dynamics | `forward_dynamics` | Image + action chunk | Video |

Pass embodiment settings through `extra_params`: `action_mode`, `domain_name` (for example `bridge_orig_lerobot`, `av`, or `camera_pose`), `raw_action_dim`, and `action_chunk_size`. Forward dynamics also takes an `action_path` pointing at an action file the server can read, so start the server with `--allowed-local-media-path` covering that file (for Docker, mount the file and pass the container-visible path). For the full set of robot, autonomous-vehicle, and camera-pose variants, see the [Cosmos 3 vLLM-Omni recipes](https://github.com/vllm-project/vllm-omni/tree/main/recipes/cosmos3).

Example video request:

```shell
curl -sS -X POST http://localhost:8000/v1/videos/sync \
  --form-string "prompt=A small warehouse robot moves a blue box across a clean floor." \
  --form-string "negative_prompt=blurry, distorted, low quality" \
  --form-string "size=1280x720" \
  --form-string "num_frames=189" \
  --form-string "fps=24" \
  --form-string "num_inference_steps=35" \
  --form-string "guidance_scale=6.0" \
  --form-string "flow_shift=10.0" \
  --form-string "seed=0" \
  --form-string 'extra_params={"use_resolution_template":false,"use_duration_template":false,"guardrails":true}' \
  -o cosmos3_t2v_output.mp4
```

Use `--form-string` for text fields (`prompt`, `negative_prompt`, `extra_params`) rather than `-F`: with `-F`, `curl` treats `;` as a content-type separator and silently truncates any value that contains one.

Common request fields (the image endpoint follows the [Image Generation API](https://docs.vllm.ai/projects/vllm-omni/en/latest/serving/image_generation_api/), and the video endpoints follow the [Videos API](https://docs.vllm.ai/projects/vllm-omni/en/latest/serving/videos_api/#request-parameters)):

| Field | Purpose |
| --- | --- |
| `prompt` | Positive text prompt |
| `negative_prompt` | Concepts or artifacts to avoid |
| `size` | Output resolution as `<width>x<height>` |
| `num_frames`, `fps` | Video length and frame rate (video endpoints only) |
| `num_inference_steps` | Diffusion denoising steps |
| `guidance_scale` | Classifier-free guidance scale (use this for Cosmos 3 CFG; do not use `true_cfg_scale`) |
| `flow_shift` | Scheduler flow-shift value |
| `seed` | Reproducibility seed |
| `max_sequence_length` | Maximum number of prompt tokens kept for conditioning (Cosmos 3 default `512`); longer prompts are truncated with a warning, shorter ones padded |
| `input_reference` | Uploaded image or video for image-to-video, video-to-video, and action requests |
| `extra_params` | JSON-encoded Cosmos 3-specific options: action settings (`action_mode`, `domain_name`, `raw_action_dim`, `action_chunk_size`, `action_path`), video-to-video conditioning (`condition_frame_indexes_vision`, `condition_video_keep`), prompt-template toggles (`use_resolution_template`, `use_duration_template`), and the per-request `guardrails` toggle |
| `extra_args` | JSON object for Cosmos 3-specific image-endpoint options such as `use_resolution_template` |

Disabling guardrails: Cosmos 3 ships safety guardrails that screen prompts and blur faces in generated output. Disable them per request by adding `guardrails: false` to `extra_params`:

```shell
curl -sS -X POST http://localhost:8000/v1/videos/sync \
  --form-string "prompt=A small warehouse robot moves a blue box across a clean floor." \
  --form-string 'extra_params={"guardrails":false,"use_resolution_template":false,"use_duration_template":false}' \
  -o cosmos3_t2v.mp4
```

To disable guardrails server-wide so the guardrail models are never loaded (per-request overrides then cannot turn them back on), pass a deploy config — a future release replaces this with a dedicated `--cosmos3-no-guardrails` flag:

```yaml
# no_guardrails.yaml
async_chunk: false
stages:
  - stage_id: 0
    max_num_seqs: 1
    enforce_eager: true
    trust_remote_code: true
    model_class_name: Cosmos3OmniDiffusersPipeline
    model_config:
      guardrails: false
      offload_guardrail_models: false
```

```shell
vllm serve nvidia/Cosmos3-Nano --omni \
  --model-class-name Cosmos3OmniDiffusersPipeline \
  --deploy-config no_guardrails.yaml \
  --port 8000 \
  --init-timeout 1800
```

References:

- [Cosmos 3 vLLM-Omni recipes](https://github.com/vllm-project/vllm-omni/tree/main/recipes/cosmos3)
- [vLLM-Omni Videos API](https://docs.vllm.ai/projects/vllm-omni/en/latest/serving/videos_api/)
- [vLLM-Omni Image Generation API](https://docs.vllm.ai/projects/vllm-omni/en/latest/serving/image_generation_api/)

</details>

#### Reasoner with Transformers
Coming soon!

#### Reasoner with vLLM

<details>
<summary>Use vLLM for Reasoner production inference behind an OpenAI-compatible chat-completions API.</summary>

```shell
uv venv --python 3.13 --seed --managed-python
source .venv/bin/activate
uv pip install --torch-backend=cu130 "vllm==0.21.0" \
  "vllm-cosmos3 @ git+https://github.com/NVIDIA/cosmos-framework.git#subdirectory=packages/vllm-cosmos3"
```

The vLLM version and the torch backend are paired: use `--torch-backend=cu130 "vllm==0.21.0"` for a CUDA 13 driver, or `--torch-backend=cu128 "vllm==0.19.1"` for CUDA 12.8. vLLM does not publish wheels for every CUDA minor version, so `--torch-backend=auto` is not reliable here — pick the pair that matches your driver.

```shell
vllm serve nvidia/Cosmos3-Nano \
  --hf-overrides '{"architectures": ["Cosmos3ReasonerForConditionalGeneration"]}' \
  --async-scheduling \
  --allowed-local-media-path / \
  --port 8000
```

For notebook launch commands (Cosmos3-Super on four GPUs, media-path defaults, and
full flag sets), see
[cookbooks/cosmos3/README.md — Start the server](cookbooks/cosmos3/README.md#start-the-server).

If your vLLM build reports that DeepGEMM is unavailable, disable it before starting the server:

```shell
export VLLM_USE_DEEP_GEMM=0
```

Configuration notes:

| Option | Use |
| --- | --- |
| `--tensor-parallel-size` | Number of GPUs used for tensor parallel inference |
| `--mm-encoder-tp-mode data` | Data parallelism for the visual encoder in multimodal workloads |
| `--media-io-kwargs '{"video": {"num_frames": -1}}'` | Allows the processor to consider all available frames before downstream frame sampling |
| `--allowed-local-media-path` | Required when requests pass local `file://` media paths |
</details>

#### Reasoner with NIM

<details>
<summary>Expand NIM Reasoner setup, container launch, and request reference</summary>

Use the [Cosmos 3 Reasoner NIM](https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/cosmos3-reasoner) for the fastest path to a production-grade, OpenAI-compatible Reasoner endpoint. NIM ships a prebuilt, optimized container so you skip the vLLM dependency and CUDA-pairing setup above; it serves text outputs from text, image, and video inputs.

You can try it interactively in your browser on the [cosmos3-nano-reasoner build page](https://build.nvidia.com/nvidia/cosmos3-nano-reasoner) — that playground is powered by this same NIM. See the [Cosmos Reason 3 NIM API reference](https://docs.nvidia.com/nim/vision-language-models/1.7.0/examples/cosmos-reason3/api.html) for the full request reference.

The container serves two sizes, selected with `NIM_MODEL_SIZE`:

| `NIM_MODEL_SIZE` | Served model name |
| --- | --- |
| `nano` (default) | `nvidia/cosmos3-nano-reasoner` |
| `super` | `nvidia/cosmos3-super-reasoner` |

Launch the NIM container (Nano shown; set `-e NIM_MODEL_SIZE=super` for Super). `NGC_API_KEY` must be set in your environment first — generate one from [NGC](https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/cosmos3-reasoner) and log Docker in to `nvcr.io` once (`docker login nvcr.io`, username `$oauthtoken`, password = your key).

```shell
export CONTAINER_NAME="nvidia-cosmos3-reasoner"
export IMG_NAME="nvcr.io/nim/nvidia/cosmos3-reasoner:1.7.0"
export LOCAL_NIM_CACHE=~/.cache/nim
mkdir -p "$LOCAL_NIM_CACHE"

docker run -it --rm --name=$CONTAINER_NAME \
  --runtime=nvidia \
  --gpus all \
  --shm-size=32GB \
  -e NGC_API_KEY=$NGC_API_KEY \
  -e NIM_MODEL_SIZE=nano \
  -v "$LOCAL_NIM_CACHE:/opt/nim/.cache" \
  -u $(id -u) \
  -p 8000:8000 \
  $IMG_NAME
```

The OpenAI-compatible API is then available at `http://127.0.0.1:8000/v1`. Query it with `curl`:

```shell
curl -X POST 'http://127.0.0.1:8000/v1/chat/completions' \
  -H 'Accept: application/json' \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "nvidia/cosmos3-nano-reasoner",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": [
        {"type": "image_url", "image_url": {"url": "https://github.com/nvidia-cosmos/cosmos-dependencies/raw/refs/heads/assets/cosmos3/inputs/vision/robot_153.jpg"}},
        {"type": "text", "text": "Describe what is happening in this image in one sentence."}
      ]}
    ],
    "max_tokens": 256,
    "stream": false
  }'
```

Or with the OpenAI Python client:

```python
from openai import OpenAI

# The container exposes the OpenAI-compatible API locally; the api_key is unused.
client = OpenAI(base_url="http://127.0.0.1:8000/v1", api_key="not-used")

response = client.chat.completions.create(
    model="nvidia/cosmos3-nano-reasoner",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {
            "role": "user",
            "content": [
                {"type": "video_url", "video_url": {"url": "https://download.samplelib.com/mp4/sample-5s.mp4"}},
                {"type": "text", "text": "List the notable events with approximate timestamps."},
            ],
        },
    ],
    max_tokens=256,
    stream=False,
    extra_body={"media_io_kwargs": {"video": {"fps": 4.0}}},
)
print(response.choices[0].message.content)
```

Inputs and request notes:

| Item | Notes |
| --- | --- |
| Images | JPG/JPEG/PNG, passed as a public URL or base64 data URI via `image_url` |
| Videos | MP4, passed as a public URL, base64, or pre-decoded `video_frames` via `video_url` |
| `extra_body.media_io_kwargs` | Controls video frame sampling, e.g. `{"video": {"fps": 4.0}}` or `{"video": {"num_frames": 16}}` |
| `extra_body.mm_processor_kwargs` | Per-image resize bounds, e.g. `{"size": {"shortest_edge": 1568, "longest_edge": 262144}}` (defaults: `shortest_edge=3136`, `longest_edge=12845056`) |
| Explicit reasoning | Append `Answer the question in the following format: <think>\nyour reasoning\n</think>\n\n<answer>\nyour answer\n</answer>.` to the user prompt |

References:

- [Cosmos Reason 3 NIM API reference](https://docs.nvidia.com/nim/vision-language-models/1.7.0/examples/cosmos-reason3/api.html)
- [cosmos3-reasoner NGC container](https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/cosmos3-reasoner)
- [cosmos3-nano-reasoner build page](https://build.nvidia.com/nvidia/cosmos3-nano-reasoner)

</details>

### Troubleshooting

#### Which CUDA version should I use?

CUDA 13 (recommended) or 12.8. Your system CUDA and PyTorch's CUDA major version must match — check with `nvidia-smi` and `python -c "import torch; print(torch.version.cuda)"`.

#### Which base container should I use?

NVIDIA NGC PyTorch: `nvcr.io/nvidia/pytorch:25.09-py3` for CUDA 13, or `nvcr.io/nvidia/pytorch:25.06-py3` for CUDA 12.

#### `torch.cuda.is_available()` is `False` ("The NVIDIA driver on your system is too old")

The installed `torch` is newer CUDA than your driver — `uv pip install torch` defaults to CUDA 13 (`cu130`). Install a matching build: `uv pip install --torch-backend=auto torch torchvision` (or pin, e.g. `--torch-backend=cu128`). For `uv sync` notebooks use `COSMOS3_UV_GROUP=cu128-train`; for vLLM pair `cu128` with `vllm==0.19.1`.

#### Import fails with `libxcb.so.1: cannot open shared object file`

On headless servers and minimal containers, importing or running a pipeline can fail with `libxcb.so.1: cannot open shared object file` (or another missing graphics library), because a dependency links against system X11/graphics libraries that are not installed. Install them:

```shell
apt-get install -y libxcb1 libgl1 libglib2.0-0
```

#### `uv` errors on install or `sync`

The Cosmos Framework requires `uv >= 0.11.3` (enforced via its `pyproject.toml`). Older versions fail to parse the project config (for example the `[tool.uv.audit]` section) and do not recognize newer `--torch-backend` values such as `cu130` (you may see `a value is required for '--torch-backend'` or an accepted-values list that stops at `cu129`). Upgrade with `uv self update` (or reinstall from <https://astral.sh/uv>).

### Choosing an Integration

| Goal | Use | Notes |
| --- | --- | --- |
| Generator research or model development | Diffusers | Python-first path for inspecting and modifying generator behavior |
| Generator production inference | vLLM-Omni | API path for image, video, sound, and action outputs |
| Reasoner research or model development | Transformers (coming soon) | Python-first path for prompts, processors, and model behavior |
| Reasoner production inference | vLLM | OpenAI-compatible endpoint for text outputs from text and vision inputs |
| Reasoner turnkey deployment | NIM | Prebuilt, optimized OpenAI-compatible container — no vLLM/CUDA setup |
| Runnable setup, training, or evaluation | Cosmos Framework | Full workflow docs for setup, inference, omni-model training, and evaluation |

### Examples

We are building examples that show Cosmos 3 capabilities end to end, including world generation, world understanding, captioning, temporal localization, grounding, and physical reasoning. Each example is a self-contained script or notebook you can run from this repository.

| Example | Surface | Workflows demonstrated | Open | nbviewer |
| --- | --- | --- | --- | --- |
| Generator (audiovisual) with Diffusers | Generator | Text-to-image, plus text-to-video and image-to-video each with or without synchronized sound, via `Cosmos3OmniPipeline`. | [Notebook](cookbooks/cosmos3/generator/audiovisual/run_with_diffusers.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/generator/audiovisual/run_with_diffusers.ipynb) |
| Generator (audiovisual) with Cosmos Framework | Generator | Text-to-image, plus text-to-video and image-to-video each with sound on or off, through the `cosmos_framework.scripts.inference` entrypoint. | [Notebook](cookbooks/cosmos3/generator/audiovisual/run_with_cosmos_framework.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/generator/audiovisual/run_with_cosmos_framework.ipynb) |
| Generator (audiovisual) with vLLM-Omni | Generator | Text-to-image, plus text-to-video and image-to-video each with sound on or off, against an OpenAI-compatible vLLM-Omni server. | [Notebook](cookbooks/cosmos3/generator/audiovisual/run_with_vllm_omni.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/generator/audiovisual/run_with_vllm_omni.ipynb) |
| Forward dynamics with Cosmos Framework | Generator | Forward dynamics: action-conditioned future-observation prediction for AV, DROID, and UMI, through the `cosmos_framework.scripts.inference` entrypoint. | [Notebook](cookbooks/cosmos3/generator/action/run_fd_with_cosmos_framework.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/generator/action/run_fd_with_cosmos_framework.ipynb) |
| Forward dynamics with vLLM-Omni | Generator | Forward dynamics: action-conditioned future-observation prediction for AV, DROID, and UMI, against an OpenAI-compatible vLLM-Omni server. | [Notebook](cookbooks/cosmos3/generator/action/run_fd_with_vllm.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/generator/action/run_fd_with_vllm.ipynb) |
| Inverse dynamics with Cosmos Framework | Generator | Inverse dynamics: ego-motion trajectory prediction from input AV video, through the `cosmos_framework.scripts.inference` entrypoint. | [Notebook](cookbooks/cosmos3/generator/action/run_id_with_cosmos_framework.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/generator/action/run_id_with_cosmos_framework.ipynb) |
| Inverse dynamics with vLLM-Omni | Generator | Inverse dynamics: ego-motion trajectory prediction from input AV video, against an OpenAI-compatible vLLM-Omni server. | [Notebook](cookbooks/cosmos3/generator/action/run_id_with_vllm.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/generator/action/run_id_with_vllm.ipynb) |
| Reasoner with Cosmos Framework | Reasoner | Text and image reasoning: detailed captioning, robot task planning, 2D grounding, describe-anything, and action-trajectory prompts, through the `cosmos_framework.scripts.inference` entrypoint. | [Notebook](cookbooks/cosmos3/reasoner/run_with_cosmos_framework.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/reasoner/run_with_cosmos_framework.ipynb) |
| Reasoner with vLLM | Reasoner | Image and video reasoning: captioning, temporal localization, embodied reasoning, common-sense reasoning, 2D grounding, describe-anything, action CoT, driving scenes, physical-plausibility, and situation understanding, against an OpenAI-compatible vLLM server (Cosmos3-Super on 4 GPUs by default; switch to Nano per the cookbook README). | [Notebook](cookbooks/cosmos3/reasoner/run_with_vllm.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/reasoner/run_with_vllm.ipynb) |
| Reasoner with NIM | Reasoner | The same image and video reasoning examples as the vLLM notebook, run against the prebuilt, OpenAI-compatible [Cosmos 3 Reasoner NIM](https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/cosmos3-reasoner) container; local media is sent as base64 data URIs. | [Notebook](cookbooks/cosmos3/reasoner/run_with_nim.ipynb) | [![Render with nbviewer](https://raw.githubusercontent.com/jupyter/design/master/logos/Badges/nbviewer_badge.svg)](https://nbviewer.org/github/nvidia/cosmos/blob/main/cookbooks/cosmos3/reasoner/run_with_nim.ipynb) |

### Inference Benchmarks

Cosmos 3 latency and serving numbers live in [`inference_benchmarks.md`](inference_benchmarks.md). Generator sections report diffusion-path latency (seconds) by GPU, engine, resolution, and tensor-parallel width; the Reasoner section reports vLLM serving metrics under concurrent load. Empty cells mean a combination has not been measured yet, not that it is unsupported.

| Benchmark | Surface | Model | What it covers |
| --- | --- | --- | --- |
| [Cosmos3-Nano generator](inference_benchmarks.md#cosmos3-nano-generator) | Generator | Cosmos3-Nano | Text-to-image, text-to-video, and image-to-video latency across PyTorch, vLLM-Omni, and Diffusers |
| [Cosmos3-Super generator](inference_benchmarks.md#cosmos3-super-generator) | Generator | Cosmos3-Super | The same modalities and engines at the larger checkpoint scale |
| [Cosmos3-Nano reasoner](inference_benchmarks.md#cosmos3-nano-reasoner) | Reasoner | Cosmos3-Nano | vLLM serving metrics — TTFT, request latency, and throughput at concurrency 1/64/128/256 |

### Finetune

Finetune Cosmos 3 with the [Cosmos Framework](https://github.com/NVIDIA/cosmos-framework), NVIDIA's end-to-end Physical AI framework for training and serving world models. It provides runnable setup, inference, omni-model training, and evaluation workflows for the Generator and Reasoner surfaces, with reference recipes for vision, action, and reasoning post-training.

See the [Cosmos Framework training guide](https://github.com/NVIDIA/cosmos-framework/blob/main/docs/training.md) for the full post-training workflow, including data preparation, configuration, and launch commands.

### Limitations

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. Applications that require physically grounded simulation, safety-critical control, or complex multi-agent behavior need additional validation, guardrails, and system-level safety analysis before deployment.

## Ecosystem

| Project | Purpose |
| --- | --- |
| [Cosmos Framework](https://github.com/NVIDIA/cosmos-framework) | End-to-end Physical AI framework for training and serving world models, including setup, inference, and training |
| [Cosmos Curator](https://github.com/NVIDIA/cosmos-curator) | Distributed Physical AI data curation system covering processing, annotation, filtering, and deduplication |
| [Cosmos Evaluator](https://github.com/NVIDIA/cosmos-evaluator) | Automated Physical AI evaluation system for world generation and world reasoning outputs |

## News

* [May 31, 2026] We released Cosmos 3 in the [NVIDIA Cosmos 3 Hugging Face collection](https://huggingface.co/collections/nvidia/cosmos3) and [Cosmos Framework](https://github.com/NVIDIA/cosmos-framework) which provides runnable setup, inference, training, and evaluation workflows for Cosmos models. See the [Cosmos 3 Technical Report](https://research.nvidia.com/labs/cosmos-lab/cosmos3/technical-report.pdf).

## License and Contact

This project may download and install additional third-party open source software projects. Review the license terms of those projects before use.

NVIDIA Cosmos source code and models are released under, and subject to the terms of, the [OpenMDW-1.1](https://openmdw.ai/license/1-1/) License. For a custom license, contact [cosmos-license@nvidia.com](mailto:cosmos-license@nvidia.com).