---
title: MiniMax-H3 Video+Audio on 8x MI300X
description: MiniMax-H3 text-to-video with synchronised audio, served by SGLang on ROCm across 8x AMD Instinct MI300X. Both FL2VA and Ref2VA partitions are baked …
product: Marketplace
url: https://docs.digitalocean.com/products/marketplace/catalog/minimax-h3-video-audio-on-8x-mi300x/
last_updated: "2026-08-06"
---

> **For AI agents:** The documentation index is at [https://docs.digitalocean.com/llms.txt](https://docs.digitalocean.com/llms.txt). Markdown versions of pages use the same URL with `index.html.md` in place of the HTML page (for example, append `index.html.md` to the directory path instead of opening the HTML document).

# MiniMax-H3 Video+Audio on 8x MI300X

Generated on 6 Aug 2026 from [the MiniMax-H3 Video+Audio on 8x MI300X catalog page](https://marketplace.digitalocean.com/apps/minimax-h3-video-audio-on-8x-mi300x)

## MiniMax-H3 on 8x MI300X

This 1-Click image serves **MiniMax-H3**, a diffusion model that generates a video clip **and its matching audio track together in a single pass**, on one 8x AMD Instinct MI300X node. Inference runs on **SGLang in diffusion mode** across all 8 GPUs with 8-way sequence parallelism and the **aiter** attention backend tuned for gfx942.

### Both partitions are baked in

The image carries **403 GB of weights** under `/opt/models/MiniMax-H3`, including **both** released partitions:

- **FL2VA** for text-to-video with audio
- **Ref2VA** for reference-image-to-video with audio

There is no first-boot download and no runtime dependency on Hugging Face being reachable. The build fails loudly if either partition is missing, so a published image always carries the full payload.

### A browser studio, not just an API

Boot the droplet and you get an **AMD-branded web studio**: type a prompt, press generate, watch progress, and download the finished MP4. No client libraries and no notebook required. The studio listens on loopback only and is reached with a one line SSH port-forward printed in the login banner, so it is never exposed to the internet by default. An OpenAI-style HTTP API is available on the same host for programmatic use.

### The gfx942 correctness fix

On gfx942 the default scaled-dot-product-attention backend produces a repeating lattice artifact in the decoded video. The failure is **silent**: the job reports success and the MP4 contains valid H.264 and AAC streams, so mechanical checks all pass while the picture is corrupted. This image pins `MINIMAX_H3_TORCH_SDPA_BACKEND=math`, which produces correct output, and the studio’s health endpoint reports the pin so you can confirm it is in force.

### Secured on first boot

The studio is protected by an **API key generated uniquely on that droplet at first boot**. No key is baked into the image, so nothing secret is published. The inference server itself binds to loopback and is never exposed directly.

### 8x only

The snapshot carries a large `min_disk` because of the baked weights, so DigitalOcean will only restore it onto the 8x MI300X SKU. That is also the only configuration with enough VRAM to shard this model.

## Software Included

| Package | Version | License |
|---|---|---|
| [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) | FL2VA+Ref2VA | other |
| [SGLang (diffusion)](https://github.com/sgl-project/sglang) | gfx942 aiter build | Apache-2.0 |
| [ROCm](https://rocm.docs.amd.com) | 7.0 | MIT/Apache-2.0 |

## Deploying this Offering using the Control Panel

Click the **Deploy to DigitalOcean** button to deploy this offering. If you aren’t logged in, this link will prompt you to log in with your DigitalOcean account.

\[![Deploy MiniMax-H3 Video+Audio on 8x MI300X to DO](https://www.deploytodo.com/do-btn-blue.svg)]([https://cloud.digitalocean.com/gpus/new?appId=3e77cfb6475958b0b04fdcb8&image=MiniMax-H3](https://cloud.digitalocean.com/gpus/new?appId=3e77cfb6475958b0b04fdcb8&image=MiniMax-H3) Video+Audio on 8x MI300X FL2VA+Ref2VA on Ubuntu 24.04&type=applications)

## Getting Started After Deploying MiniMax-H3 Video+Audio on 8x MI300X

After the droplet boots, SSH in as root. The login banner (MOTD) is the primary interface: it reports whether the model has finished loading, prints the exact command to open the studio, and prints the API key generated for this specific droplet. Nothing is pre-shared, the key is created on first boot.

The studio binds to loopback and is never exposed to the internet, so reach it with an SSH port-forward from your own machine:

```
ssh -N -L 8080:127.0.0.1:8080 root@YOUR_DROPLET_IP
```

Then open [http://127.0.0.1:8080](http://127.0.0.1:8080), paste the API key once when prompted, type a prompt and press Generate. A 6 second clip takes roughly 4 to 5 minutes on 8x MI300X. The finished video plays in the page and can be downloaded directly.

Prompting tip: H3 takes audio direction from the prompt text. Append an `overall_soundscape:` clause describing what should be heard, for example `a wide shot of a desert at sunrise. overall_soundscape: wind and distant birds`. Without it you still get an audio track, but it is unguided.

To use the HTTP API instead, POST to `http://localhost:30010/v1/videos` with a JSON body containing `model`, `prompt` and `duration`, then poll `/v1/videos/<id>` until status is completed.

The server loads the FL2VA (text-to-video) partition by default. To serve Ref2VA (reference-image-to-video) instead, set `H3_PARTITION=Ref2VA` via `systemctl edit h3-server` and restart the service. Useful commands: `systemctl status h3-server`, `journalctl -fu h3-server`, `curl -s localhost:8080/api/health`, `amd-smi monitor`.