> 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/minimax-h3.md).

# MiniMax-H3: How to Run Locally

Run MiniMax-H3 locally for video with sound, and train a MiniMax-H3 video LoRA in Unsloth.

**MiniMax-H3** is MiniMax's open video model. It makes a video and its soundtrack together from a text prompt, with optional start and end frames or reference pictures, videos and audio. Run it in [Unsloth](https://unsloth.ai/docs/desktop.md) with the Unsloth FP8 repo, and train your own MiniMax-H3 LoRA from short clips with sound.

{% columns %}
{% column width="50%" %}

* Clips are 24 fps and 5.2 to 14.4 s long.
* Default: **20 steps**, **864 × 480**, 5.2 s.
* One 864 × 480 video takes about **30 s** on an RTX PRO 6000.
* LoRA training: about **36 min** for 400 steps, 49.7 GB peak VRAM.

MiniMax-H3: [FP8](https://huggingface.co/unsloth/MiniMax-H3-FP8) • [GGUF](https://huggingface.co/unsloth/MiniMax-H3-GGUF) • [Original](https://huggingface.co/MiniMaxAI/MiniMax-H3)

Tested with Unsloth 2026.11.4 and a build from main on an RTX PRO 6000 (96 GB).
{% endcolumn %}

{% column width="50%" %}

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2Fyc83EeZRTERrxk24F75C%2Fminimax-h3-out-forest.webp?alt=media" alt="Aerial view over a misty pine forest and a winding river at sunrise"><figcaption><p>FP8 repo, 864 × 480, 5.2 s (no sound in preview)</p></figcaption></figure>
{% endcolumn %}
{% endcolumns %}

#### MiniMax-H3 LoRA finetuning - Before and after

[Fine-tune MiniMax-H3: before and after examples](https://unsloth.ai/docs/models/minimax-h3/fine-tune-minimax-h3.md#minimax-h3-lora-finetuning-before-and-after)

Unsloth also allows you to finetune MiniMax-H3! For example with 864 × 480, 5.2 s, 20 steps, seed 42, both on the original bf16 model in Unsloth. Prompt: "bbbstyle, a sunny meadow with a big oak tree and a small stream, birds chirping" we show before and after finetuning's style:

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FY40HyPvTZusQH0RDXU4I%2Fminimax-h3-lora-studio-before.webp?alt=media" alt="Without the LoRA: a realistic meadow with a big oak tree and a small stream under a blue sky"><figcaption><p>Without LoRA</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FQxzr4zJRmMzlFE6OHCDi%2Fminimax-h3-lora-studio-after.webp?alt=media" alt="With the LoRA: the same meadow in a 3D animated film look, with a stylized oak tree, bright green grass and a small stream"><figcaption><p>With bbbstyle LoRA, 200 steps</p></figcaption></figure></div>

### Memory requirements

| Task               | Download                                                                    | Total        | GPU memory                       |
| ------------------ | --------------------------------------------------------------------------- | ------------ | -------------------------------- |
| **Run** (Fast FP8) | INT8 video model 20.3 GB, INT8 Qwen3-VL 32B text encoder 27.1 GB, VAEs      | about 50 GB  | 54 GB in use                     |
| **Train a LoRA**   | Original bf16 repo: video model 66.3 GB, text encoder 66.7 GB, VAEs 11.1 GB | about 144 GB | 49.7 GB peak, 80 GB+ recommended |

Training loads the 63 GB text encoder once to encode captions, then frees it before the video model loads.

### Quickstart

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FWti8xb1Y3sOQBfG8oXMd%2Fminimax-h3-flow.webp?alt=media" alt="Animation: pick MiniMax H3 (Fast FP8), choose Text and frames, type a prompt, click Generate and play the forest clip" width="563"><figcaption><p>The Unsloth flow, sped up</p></figcaption></figure>

{% stepper %}
{% step %}

#### Install Unsloth

Use the [Unsloth Desktop app](https://unsloth.ai/docs/desktop.md) or install from the terminal.

<a href="https://unsloth.ai/download" class="button primary" data-icon="down-to-bracket">Download Unsloth</a>

MacOS, Linux, WSL:

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

Windows PowerShell:

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

Launch Unsloth and open `http://localhost:8888`:

```bash
unsloth studio -H 0.0.0.0 -p 8888
```

{% endstep %}

{% step %}

#### Open Video

Open **Images** in the sidebar and click **Video** at the top right, or go to `http://localhost:8888/video`.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FqCq3GkDxVruGcu5LGHPi%2Fminimax-h3-video-page.png?alt=media" alt="The Unsloth Video page with the Create tab open and no model loaded"><figcaption><p>The Video page</p></figcaption></figure>
{% endstep %}

{% step %}

#### Pick MiniMax-H3

Click **Select video model**, search `MiniMax-H3` and pick **MiniMax H3 (Fast FP8)**.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FbRd7iMQUQKMxQt8erID0%2Fminimax-h3-picker.png?alt=media" alt="Model picker search results for MiniMax-H3 with the Fast FP8, GGUF, FP8 and original entries" width="540"><figcaption><p>Model picker</p></figcaption></figure>

Choose a mode:

* **Text and frames**: prompt, with optional start and end frames.
* **References**: reference pictures, videos and audio. Separate model download.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FG589ovJjwsvgaLILbmrH%2Fminimax-h3-mode.png?alt=media" alt="Dialog: Choose how MiniMax H3 should generate, with the Text and frames and References options" width="450"><figcaption><p>Mode dialog</p></figcaption></figure>
{% endstep %}

{% step %}

#### Set the prompt and settings

Describe the scene, camera and sounds. Pick a size under **Trained sizes** and a **Duration**. Unsloth fills in the other settings.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FA3ocFMljlYqyFZIVQnGg%2Fminimax-h3-settings.png?alt=media" alt="Create panel with the forest prompt, empty start and end frames, 864 by 480, 5.2 seconds and 124 frames, 24 fps, 20 steps, motion shift 12, audio shift 3 and seed 42" width="400"><figcaption><p>Create settings</p></figcaption></figure>
{% endstep %}

{% step %}

#### Generate

Click **Generate**. The video plays with sound in the preview. Use **Recipe** for the settings and **Download** to save it.

<figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2F1kKizKGOjO83IGdw1PAa%2Fminimax-h3-result.png?alt=media" alt="Finished video of a misty pine forest and river at sunrise in the Video page preview, with the gallery below"><figcaption><p>Result, seed 42</p></figcaption></figure>

The first video compiles the model and takes a few minutes.
{% endstep %}
{% endstepper %}

Prompt: "A slow aerial shot glides over a misty pine forest and a winding river at sunrise. Golden light breaks through the fog, birds sing and water rushes over the rocks. Cinematic nature film."

### Start frame, end frame and references

* **Start frame**: the video begins on your picture. Drop it in **Start frame**.
* **End frame**: the video ends on your picture. Use both to move between two pictures.
* **References**: load MiniMax-H3 in **References** mode to guide the video with pictures, videos and audio tracks.
* Pictures must be between 1:4 and 4:1. Describe the motion between frames in the prompt.

### Recommended settings

| Setting          | Value                                                            |
| ---------------- | ---------------------------------------------------------------- |
| **Steps**        | 20                                                               |
| **Guidance**     | Not used (guidance-distilled)                                    |
| **Frame rate**   | 24 fps                                                           |
| **Duration**     | 5.2 s (124 frames) to 14.4 s (345 frames), in steps of 17 frames |
| **Size**         | 864 × 480 to start; up to 1344 × 768 or 1536 × 672               |
| **Motion shift** | 12                                                               |
| **Audio shift**  | 3                                                                |
| **Seed**         | Any number to repeat a result                                    |

#### T4 and other GPUs without bf16

Use the **GGUF** entry. Without bf16, the FP8 repo and the original weights fall back to float32 and need about 288 GB of system RAM, so on a smaller machine Unsloth refuses them before downloading anything and points you to the GGUF build (needs Unsloth newer than 2026.11.4).

On a 15 GB Colab T4, GGUF **UD-Q2\_K\_XL** with the Q2\_K\_M text encoder rendered 864 x 480, 124 frames, 20 steps in about **34 min** per clip, using about 26 GB of system RAM.

### Train a LoRA

See [Fine-tune MiniMax-H3](https://unsloth.ai/docs/models/minimax-h3/fine-tune-minimax-h3.md) for how to finetune MiniMax-H3. As an example, see how for both videos, they both took about 10 min on an RTX PRO 6000, including loading. The same pair with the 400-step LoRA is on the right.

<div><figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FQ9SnobeErT5EubX2JC3I%2Fminimax-h3-lora-meadow-before.webp?alt=media" alt="Without the LoRA: a realistic meadow with a big oak tree and a small stream under a blue sky"><figcaption><p>Without LoRA</p></figcaption></figure> <figure><img src="https://3215535692-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FxhOjnexMCB3dmuQFQ2Zq%2Fuploads%2FLrEdDhkeBntupm1TbjDb%2Fminimax-h3-lora-meadow-after.webp?alt=media" alt="With the LoRA: the same meadow in a 3D animated film look, with stylized oak trees, soft light and a small animated character"><figcaption><p>With bbbstyle LoRA, 400 steps</p></figcaption></figure></div>

#### Use the LoRA

See [Fine-tune MiniMax-H3](https://unsloth.ai/docs/models/minimax-h3/fine-tune-minimax-h3.md#use-the-lora)

#### Outside Unsloth: diffusers

See [Fine-tune MiniMax-H3](https://unsloth.ai/docs/models/minimax-h3/fine-tune-minimax-h3.md#outside-unsloth-diffusers)

### Troubleshooting

See [Fine-tune MiniMax-H3](https://unsloth.ai/docs/models/minimax-h3/fine-tune-minimax-h3.md#troubleshooting)

<details>

<summary>MiniMax-H3 is not in the Video model picker</summary>

[Update Unsloth](https://unsloth.ai/docs/get-started/install/updating) and restart it. If a model is loaded, click **Recommended** in the picker.

</details>


---

# 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/minimax-h3.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.
