> For the complete documentation index, see [llms.txt](https://unsloth.ai/docs/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://unsloth.ai/docs/models/cosmos-3.md).

# NVIDIA Cosmos 3: How to Run Locally

Run NVIDIA Cosmos 3 world models locally with Unsloth: text-to-video, image-to-video, video continuation and video restyle.

**NVIDIA Cosmos 3** is a family of omnimodal world models for Physical AI. **Cosmos3-Nano** (16B) and **Cosmos3-Edge** (4B) run in Unsloth for **text-to-video** and **image-to-video**, and Cosmos3-Nano can also **continue a clip**. **Cosmos3-Nano-Sim-Transfer** restyles an existing video from its edges in the **Restyle** tab. Loading and LoRA setup for the **reasoner** work with Unsloth; reasoner training is described below as an unvalidated example.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FUUdAGmDpMwHPVOpzFcp4%2Fcosmos-3-nano-t2v.gif?alt=media" alt="Cosmos3-Nano text-to-video: a robot arm places a red box on a pallet beside a conveyor" width="563"><figcaption><p>Cosmos3-Nano, 832×480, 93 frames, seed 42, about 50 s</p></figcaption></figure>

Cosmos 3 exposes two runtime surfaces. NVIDIA’s **Generator** model can take text, vision, sound and action as input and produce video, sound and action. Unsloth generates video only: it does not load the sound tokenizer, so clips have no audio track, and action input and output are not exposed. The **Reasoner** takes text and vision and produces text for world understanding, physical reasoning, task planning and action forecasting.

Text-to-video, image-to-video and continuation run locally through the Cosmos3OmniPipeline diffusers pipeline. Restyle runs Cosmos3-Nano-Sim-Transfer on stable-diffusion.cpp.

* **Cosmos3-Edge (4B):** [Hugging Face](https://huggingface.co/nvidia/Cosmos3-Edge)
* **Cosmos3-Nano (16B):** [Hugging Face](https://huggingface.co/nvidia/Cosmos3-Nano)
* **Cosmos3-Nano-Sim-Transfer:** [Hugging Face](https://huggingface.co/nvidia/Cosmos3-Nano-Sim-Transfer)
* **Docs:** [NVIDIA Cosmos documentation](https://docs.nvidia.com/cosmos/latest/cosmos3/model_reference.html) and the [Cosmos repository](https://github.com/NVIDIA/cosmos)

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FwOasrmOzKWt7wJQLju9g%2Fcosmos-3-flow.gif?alt=media" alt="Animation: search Cosmos3 in the Video model picker, pick Cosmos3 Edge, wait for the download and load, add a start frame, click Generate and get a robot arm clip" width="563"><figcaption><p>The Unsloth flow with Cosmos3-Edge, sped up</p></figcaption></figure>

{% hint style="info" %}
**Note**

Cosmos 3 generator support landed in [PR #14554](https://github.com/unslothai/unsloth/pull/14554). Use Unsloth v0.1.917-beta or newer for 256p generation, model-specific frame rates, Cosmos3-Nano in the model picker, Nano video continuation and Restyle. The release after v0.1.918-beta adds the changes from [PR #18116](https://github.com/unslothai/unsloth/pull/18116): the safety checker off by default (Step 2), the **Set up Restyle** button (Step 6), **Low VRAM** that never uses more memory than Balanced, and the Video page opening at the trained 832×480. See [Updating Unsloth](https://unsloth.ai/docs/get-started/install/updating).
{% endhint %}

## Memory and speed

Measured in Unsloth (main, November 16) on an NVIDIA RTX PRO 6000 (96 GB), one clip at a time, seed 42, model defaults (Edge 20 steps, Nano 35 steps, guidance 6, 24 FPS). Peak VRAM is the whole GPU as reported by `nvidia-smi`, read in a freshly started. These are measured configurations, not minimum hardware requirements.

| Model                         | Mode                         | Memory                    | Settings                    | Time    | Peak VRAM    |
| ----------------------------- | ---------------------------- | ------------------------- | --------------------------- | ------- | ------------ |
| **Cosmos3-Edge (4B)**         | Text-to-video                | Auto (resident)           | 832×480, 53 frames          | 6.8 s   | 13.6 GiB     |
| **Cosmos3-Edge (4B)**         | Text-to-video                | Auto (resident)           | 832×480, 121 frames         | 16.6 s  | 14.0 GiB     |
| **Cosmos3-Edge (4B)**         | Text-to-video                | Auto (resident)           | 832×480, 149 frames         | 20.4 s  | 13.7 GiB     |
| **Cosmos3-Edge (4B)**         | Image-to-video               | Auto (resident)           | 832×480, 121 frames         | 15.9 s  | about 15 GiB |
| **Cosmos3-Edge (4B)**         | Text-to-video                | Balanced (INT8, streamed) | 832×480, 121 frames         | 23.5 s  | 7.6 GiB      |
| **Cosmos3-Nano (16B)**        | Text-to-video                | Auto (resident)           | 832×480, 93 frames          | 49.8 s  | 36.4 GiB     |
| **Cosmos3-Nano (16B)**        | Text-to-video                | Auto (resident)           | 832×480, 189 frames         | 118.8 s | 36.8 GiB     |
| **Cosmos3-Nano (16B)**        | Text-to-video                | Auto (resident)           | 1280×720, 189 frames        | 394.9 s | 44.8 GiB     |
| **Cosmos3-Nano (16B)**        | Image-to-video, continuation | Auto (resident)           | 832×480, 49 frames          | 23.8 s  | under 43 GiB |
| **Cosmos3-Nano (16B)**        | Text-to-video                | Balanced (INT8, streamed) | 832×480, 49 frames          | 41.6 s  | 9.0 GiB      |
| **Cosmos3-Nano-Sim-Transfer** | Restyle (BF16)               | sd.cpp                    | 832×480, 121 frames, 30 FPS | 37.5 s  | 39.8 GiB     |
| **Cosmos3-Nano-Sim-Transfer** | Restyle (Q8\_0 GGUF)         | sd.cpp                    | 832×480, 121 frames, 30 FPS | 37.2 s  | 29.0 GiB     |

Loading from the local cache took about 8 seconds for either model on this machine; the first download took about 1 minute for Edge (9.2 GB) and about 3 minutes for Nano (about 33 GB, the sound tokenizer is skipped). With **Memory: Auto** on a GPU where the weights do not fit, Unsloth picks an offload mode by itself. **Balanced** streams the transformer and quantizes it to INT8 on load: Edge went from 14.0 GiB to 7.6 GiB at 121 frames, and Nano used 9.0 GiB at 49 frames where its resident weights alone take about 31 GiB, for about 1.5x the time per step. On v0.1.918-beta and earlier, **Low VRAM** used more memory than Balanced for Cosmos 3 (18.9 GiB for Nano). From the next release it streams the transformer and the text encoder, so it never peaks above Balanced: on the same GPU, Nano's Low VRAM peak went from 18.9 GiB to 8.4 GiB and Edge's from 8.6 GiB to 7.6 GiB, at Balanced's speed.

NVIDIA has shared pre-quantized FP8 checkpoints for Cosmos3-Nano, Cosmos3-Super and Cosmos3-Edge. Unsloth Desktop v0.1.920-beta supports loading Nano-FP8 and Edge-FP8 on the Video page, with no extra NVIDIA package. Super-FP8 loading in Studio remains unverified. Unsloth's INT8 is applied at load, and BF16 is the default precision. With speed optimizations off, Unsloth's output matched the plain diffusers pipeline at the same seed to within video encoding noise; the default speed settings give the same scene and motion with small differences in detail.

## Supported generation settings

| Setting               | Supported values                                                                                               |
| --------------------- | -------------------------------------------------------------------------------------------------------------- |
| **Resolution tiers**  | Nano: 256p, 480p, 720p. Edge: 256p, 480p. A request without a size uses 1280×720 on Nano and 832×480 on Edge.  |
| **Aspect ratios**     | 16:9, 4:3, 1:1, 3:4, 9:16 at 480p and 720p. The 256p sizes are 320×192, 192×320, 256×256, 320×256 and 256×320. |
| **Frame rates**       | Nano: 10, 16, 24 or 30 FPS. Edge: 12–30 FPS. Default: 24 FPS for both.                                         |
| **Frame count**       | Nano: 5–397 (default 189). Edge: 53–149 (default 121). Both use frame counts of 4k+1.                          |
| **Precision**         | BF16, or INT8 applied at load by the Balanced and Low VRAM memory modes                                        |
| **GPU architectures** | Measured here on an NVIDIA RTX PRO 6000 (Blackwell). Other GPUs are not covered by these measurements.         |

Unsloth refuses settings outside these ranges with a message that names the nearest supported value, for example a 50 frame request on Edge suggests 53.

{% hint style="warning" %}
**Important**

**Cosmos3-Edge** supports 256p and 480p, 12–30 FPS, and frame counts of 53–149 in Unsloth (default 121). It does not continue clips. Nano continues a clip from five conditioning frames. Restyle uses the separate Cosmos3-Nano-Sim-Transfer model.
{% endhint %}

For 16:9 output, **vision conditioning** uses 1280×720 at 720p and 832×480 at 480p. The pipeline resizes and crops conditioning images to the output size, so matching the output aspect ratio avoids cropping. Nano video continuation keeps five conditioning frames. Keep world-generation prompts under roughly 300 words.

## Quickstart

### Step 1. Set up Unsloth

The easiest way to get started is by downloading the [Unsloth Desktop app](https://unsloth.ai/docs/desktop). Works on [macOS](https://unsloth.ai/docs/get-started/install/mac), [Windows](https://unsloth.ai/docs/get-started/install/windows-installation), and [Linux](https://unsloth.ai/docs/get-started/install/linux).

Or install manually. MacOS, Linux, WSL:

```bash
curl -fsSL https://unsloth.ai/install.sh | sh
```

Windows PowerShell:

```powershell
irm https://unsloth.ai/install.ps1 | iex
```

Unsloth installs a diffusers version that runs Cosmos 3, so nothing else is needed for Cosmos3-Edge. Unsloth refuses Cosmos 3 on diffusers older than 0.41 with a message that says how to update.

### Step 2. (Optional) Turn on the safety checker

The Cosmos 3 model repositories are not gated, and NVIDIA's content safety filter (the Cosmos guardrail) is **off by default** in Unsloth for every Cosmos 3 model, so you can skip this step. The load log says when it is off. This default needs the release after v0.1.918-beta; earlier releases turn the guardrail on for Cosmos3-Nano and refuse to load it until it is set up or `UNSLOTH_COSMOS3_SAFETY_CHECKER=0` is set.

To turn the safety checker on:

1. Start Unsloth with the safety checker on:

```bash
UNSLOTH_COSMOS3_SAFETY_CHECKER=1 unsloth studio
```

2. Install `cosmos-guardrail` into Unsloth's own environment. When you load a Cosmos 3 model without it, Unsloth shows a message with an **Install guardrail** button (for the account that owns the installation), which installs it with no restart. Or run the command the message shows; a plain `pip install` goes to a different Python, which Unsloth does not use:

```bash
uv pip install --python ~/.unsloth/studio/unsloth_studio/bin/python cosmos-guardrail
```

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FauTn508CXQokJ7s4B9Yd%2Fcosmos-3-guardrail-install.png?alt=media" alt="Message: the Cosmos safety checker needs cosmos-guardrail, with the install command, the licence link and an Install guardrail button" width="400"><figcaption><p>The safety checker is on and cosmos-guardrail is missing</p></figcaption></figure>

3. Accept the licence of [nvidia/Cosmos-1.0-Guardrail](https://huggingface.co/nvidia/Cosmos-1.0-Guardrail) with your Hugging Face account. The repository is gated, and access is approved as soon as you accept.
4. Add that account's Hugging Face token in Unsloth **Settings**. The guardrail downloads its own weights from that repository the first time a Cosmos 3 model loads.

With the safety checker on, Unsloth checks that your token can read the guardrail repository before it downloads the model. Without access, the load stops within seconds instead of after the roughly 33 GB Nano download.

### Step 3. Generate a video

Open **More → Video** in the sidebar. Click the model selector, search for **Cosmos3** (or **Cosmos3-Edge** / **Cosmos3-Nano**) and click a model. Unsloth downloads it, with progress in the Downloads panel, and loads it.

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FzV5LPOvMdCtYikhCnqrC%2Fcosmos-3-more-video.png?alt=media" alt="Sidebar with the More menu open and Video in the list" width="240"><figcaption><p>More → Video</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FPxq0J6cSeQsu6uDm6RMl%2Fcosmos-3-picker-search.png?alt=media" alt="Video model picker search results with Cosmos3 Nano and Cosmos3 Edge" width="560"><figcaption><p>Search results</p></figcaption></figure></div>

Describe the scene and click **Generate**. Start at **480p, 16:9, 24 FPS** with a low frame count and increase once you know the clip fits in memory. The Video page opens Cosmos3-Edge and Cosmos3-Nano at **832 × 480**, a trained size, and **Trained sizes** shows the size in use (on v0.1.918-beta and earlier, open **Trained sizes** and choose **832 × 480**). Then pick a duration. Use a trained size; the page warns when the size is not one of them.

Your result appears in the gallery under the preview. **Recipe** shows the clip's settings and can restore them, and **Download** saves the MP4.

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FQt4SWKQvNu2g1kT1cvZW%2Fcosmos-3-result.png?alt=media" alt="Video page with a finished Cosmos3-Edge clip of a robot arm and a box in a warehouse, and the gallery below"><figcaption><p>Cosmos3-Edge, 832×480, 121 frames, about 16 s</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FzAEkzEaTWgMGtnMo5fz6%2Fcosmos-3-recipe.png?alt=media" alt="Generation settings card with prompt, model, size 832 by 480, 121 frames at 24 fps, 20 steps, guidance 6 and seed 42" width="326"><figcaption><p>Recipe card</p></figcaption></figure></div>

### Step 4. Image-to-video

In **Create**, drop an image file on the **Start frame** box under the prompt and describe the motion you want. Cosmos 3 conditions on the image and generates forward from it; the first frame of the clip reproduces the image.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2F3BphTOcb5eLWnkPv0ksE%2Fcosmos-3-settings.png?alt=media" alt="Create settings with a start frame, the aspect ratio set to Match source, 832 by 480, 121 frames, 24 FPS, 20 steps and guidance 6" width="400"><figcaption><p>Start frame staged on Cosmos3-Edge</p></figcaption></figure>

The pipeline resizes and crops your conditioning image to the output size. Match the target aspect ratio to avoid cropping; at 16:9, 480p is 832×480 and Nano’s 720p is 1280×720.

### Step 5. Continue a clip (Nano)

With Cosmos3-Nano loaded, drop a video on **Continue a clip**, choose **From its last frames** (or its first frames), describe what happens next and click **Generate**. Nano keeps five frames of the clip and generates from there. Edge does not show this section.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FVAYRy5lq1iqte46gtG88%2Fcosmos-3-continue.png?alt=media" alt="Continue a clip with an uploaded clip and Continue set to From its last frames" width="400"><figcaption><p>Continue a clip</p></figcaption></figure>

### Step 6. Restyle a video (Cosmos3-Nano-Sim-Transfer)

The **Restyle** tab keeps a clip's shapes and motion, traced as Canny edges, and repaints it from your prompt with Cosmos3-Nano-Sim-Transfer on stable-diffusion.cpp. It uses up to the first 4 seconds at 832×480 and 30 FPS, needs an NVIDIA GPU with about 40 GB of memory (about 30 GB with a Q8\_0 GGUF), and has been tested on Linux only.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2Fw7cDhGvQh2TocbZHDhyA%2Fcosmos-3-restyle.png?alt=media" alt="Restyle tab with the input clip, its edges and the restyled clip side by side"><figcaption><p>Input, edges and result</p></figcaption></figure>

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FpftrDXCfsdtBfKbKZRv4%2Fcosmos-3-restyle-strip.gif?alt=media" alt="Animation: the input clip, its Canny edges and the restyled clip playing side by side" width="563"><figcaption><p>121 frames in 37.5 s on an RTX PRO 6000</p></figcaption></figure>

Set Restyle up once from the tab: click **Set up Restyle**. It downloads the model (about 35 GB) and needs about 67 GB of free disk while it runs and 31 GB after. Unsloth installs its stable-diffusion.cpp server, merges the transformer shards, converts the Wan2.2 VAE, and shows progress; you can cancel and resume. Only the account that owns the installation can run it. It refuses up front on macOS, without an NVIDIA GPU, or with less than 31 GB of GPU memory; a GPU with 31 to 42 GB gets a Q8\_0 copy, which uses about 11 GB less memory. Set up Restyle needs the release after v0.1.918-beta.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2Fhg78J2NonBoxrafLzovl%2Fcosmos-3-restyle-setup.png?alt=media" alt="Restyle tab: Restyle is not set up yet, with the download and disk sizes and a Set up Restyle button" width="400"><figcaption><p>Set up Restyle</p></figcaption></figure>

On v0.1.918-beta and earlier, set it up by hand:

1. Download the transformer from [nvidia/Cosmos3-Nano-Sim-Transfer](https://huggingface.co/nvidia/Cosmos3-Nano-Sim-Transfer) and merge its shards with [`scripts/convert_cosmos3.py`](https://github.com/unslothai/stable-diffusion.cpp/blob/master/scripts/convert_cosmos3.py) from unslothai/stable-diffusion.cpp:

```bash
python convert_cosmos3.py Cosmos3-Nano-Sim-Transfer/transformer cosmos3_nano_sim_transfer_bf16.safetensors
```

2. Download the Wan2.2 VAE as a single `wan2.2_vae.safetensors` file, as linked in the [stable-diffusion.cpp Cosmos 3 guide](https://github.com/unslothai/stable-diffusion.cpp/blob/master/docs/cosmos3.md).
3. Start Unsloth with both paths set:

```bash
UNSLOTH_COSMOS3_TRANSFER_MODEL=/path/to/cosmos3_nano_sim_transfer_bf16.safetensors \
UNSLOTH_COSMOS3_TRANSFER_VAE=/path/to/wan2.2_vae.safetensors \
unsloth studio
```

The manual setup also needs Unsloth's stable-diffusion.cpp server, which Unsloth downloads the first time it runs a model on stable-diffusion.cpp (for example a GGUF image model); until then the tab says it is not set up. A Q8\_0 GGUF made with `sd-cli -M convert --type q8_0` also works and uses about 11 GB less memory. In newer releases these two variables still take precedence over the files Set up Restyle manages.

## Fine-tune the Cosmos 3 Reasoner

NVIDIA reports fine-tuning the Cosmos3-Nano reasoner with Unsloth on a public robotics VQA dataset and matching its internal training stack. The example below loads Cosmos3-Edge's reasoner and attaches LoRA adapters. It ran as written with Unsloth 2026.11.5 and transformers 5.17 on an RTX PRO 6000: the model loaded with Unsloth's Cosmos3\_Edge patches, the LoRA adapters attached (25.2M trainable parameters) and the model answered a text prompt. Training and export have not been validated yet.

```python
from unsloth import FastVisionModel

model, processor = FastVisionModel.from_pretrained(
    "nvidia/Cosmos3-Edge",
    max_seq_length = 4096,
    load_in_4bit = False,
)

model = FastVisionModel.get_peft_model(
    model,
    r = 16,
    lora_alpha = 16,
    finetune_vision_layers   = True,
    finetune_language_layers = True,
)
```

Before training on your own video or image QA dataset, validate LoRA training and [export](https://unsloth.ai/docs/new/studio/export) for Cosmos 3. See [Vision Fine-tuning](https://unsloth.ai/docs/basics/vision-fine-tuning) for general dataset formatting and hyperparameter guidance; those instructions do not by themselves confirm Cosmos 3 training support.

Typical uses: robotics VQA, physical reasoning over recorded footage, task planning and action forecasting from video.

## What is supported today

| Workflow                                   | Status                                                                                                                                                               |
| ------------------------------------------ | -------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Text-to-video (Edge, Nano)                 | Supported                                                                                                                                                            |
| Image-to-video (Edge, Nano)                | Supported                                                                                                                                                            |
| Video continuation (Nano)                  | Supported                                                                                                                                                            |
| Video restyle (Cosmos3-Nano-Sim-Transfer)  | Supported; set up from the Restyle tab (manual setup on v0.1.918-beta and earlier); tested on Linux with NVIDIA GPUs                                                 |
| Reasoner loading, LoRA setup and inference | Works with the example above                                                                                                                                         |
| Reasoner LoRA training and export          | Not yet validated                                                                                                                                                    |
| Video (generator) LoRA training            | Not supported yet                                                                                                                                                    |
| Pre-quantized Cosmos 3 checkpoints         | NVIDIA has shared Nano, Super and Edge FP8 checkpoints. Nano-FP8 and Edge-FP8 load on the Video page in Desktop v0.1.920-beta; Super-FP8 loading remains unverified. |
| Cosmos3-Super (64B)                        | Not currently in Unsloth                                                                                                                                             |

## Troubleshooting

### The model will not load and asks for the guardrail

With the safety checker on (`UNSLOTH_COSMOS3_SAFETY_CHECKER=1`), click **Install guardrail** in the message (or run the command it shows), accept the licence of [nvidia/Cosmos-1.0-Guardrail](https://huggingface.co/nvidia/Cosmos-1.0-Guardrail) and add that account's Hugging Face token in Settings, as in Step 2. If the token cannot access that repository, the load stops with these steps before the model downloads; on v0.1.918-beta and earlier this check came after the download. On Unsloth releases before the default change, Cosmos3-Nano asks for the guardrail even without the variable: set it up, update Unsloth, or start Unsloth with `UNSLOTH_COSMOS3_SAFETY_CHECKER=0`.

### Unsloth refuses Cosmos 3 because of the diffusers version

Cosmos 3 needs diffusers 0.41 or newer, and Unsloth refuses older versions at load. Update Unsloth, which installs a working version.

### The model runs out of memory

Desktop v0.1.920-beta supports NVIDIA's Nano-FP8 and Edge-FP8 on the Video page; Super-FP8 loading remains unverified. To reduce memory use:

* Set **Memory** to **Balanced** in the settings. It streams the transformer in INT8: Nano used 9.0 GiB at 832×480 and 49 frames in the measurements above. From the release after v0.1.918-beta, **Low VRAM** uses no more than Balanced; on earlier releases pick Balanced.
* Drop to a lower resolution tier, 480p down to 256p.
* Reduce the frame count. Unsloth accepts 53–149 frames for Edge and 5–397 for Nano, in steps of 4.
* Use Cosmos3-Edge (4B) rather than Cosmos3-Nano (16B).

### Continue a clip is missing

Video continuation appears in **Create** only when Cosmos3-Nano is loaded. Cosmos3-Edge does not continue clips.

### Restyle says it is not set up

Click **Set up Restyle** in the tab (Step 6). If the button is missing, the message says why: another account owns the installation, the GPU has less than 31 GB, there is no NVIDIA GPU, or the OS is macOS. On v0.1.918-beta and earlier, set `UNSLOTH_COSMOS3_TRANSFER_MODEL` and `UNSLOTH_COSMOS3_TRANSFER_VAE` as in Step 6 and restart Unsloth; if the message says no stable-diffusion.cpp server is installed, run a GGUF image model on the Images page once so Unsloth downloads the server.

## FAQ

### What is the difference between the Generator and the Reasoner?

NVIDIA’s **Generator** model can produce video, sound and action. Unsloth generates video from text, images or a clip, and restyles video with Cosmos3-Nano-Sim-Transfer; sound and action generation are not exposed. The **Reasoner** produces text from text and vision; loading and LoRA setup work with Unsloth, and its training path is not yet validated.

### Which model should I start with?

Start with **Cosmos3-Edge (4B)**, the smaller and faster model, using a measured configuration in the memory table as a reference rather than a guaranteed minimum. Choose **Cosmos3-Nano (16B)** when you need 720p output or video continuation.

### Can I train the video generator, not just the reasoner?

Not yet. Unsloth does not train Cosmos 3 generator LoRAs today.

### Does Cosmos 3 generate sound?

NVIDIA’s model supports sound alongside video. Unsloth does not load the sound tokenizer, so its clips have no audio.

### Does this run fully locally?

Once the models are downloaded, generation runs on your device. With the safety checker turned on, Unsloth also downloads the guardrail's weights from Hugging Face the first time a Cosmos 3 model loads.


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://unsloth.ai/docs/models/cosmos-3.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
