MAGI-2 Preview

MAGI-2 Preview

A large audio-video generation model that makes 10-second clips with sound from a text prompt or a prompt plus a still image, refined to 1080p. For researchers with a server of eight Hopper GPUs.

Made by sand-ai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: MAGI-2 Preview is a unified audio-video generation model with 114B parameters, of which only 6B are active for each token. It is built on MagiMoE. The repository holds code for running the model. It turns a written prompt, with or without a still image, into a video that carries its own generated sound.

Who it suits: Research teams that study large video generation models and have access to a server with eight NVIDIA Hopper GPUs.

Industry | Who uses it | How they use it | Example
AI research | Video generation researchers | Study a sparse model that generates sound and picture together | Generate a 10-second clip from a written prompt
Media technology | Teams with large GPU servers | Test text-to-video and image-to-video output | A short clip with a generated audio track
Engineering | People running big inference jobs | Try the two-stage pipeline on their own hardware | A low-resolution pass followed by a 1080p refinement

You can skip it if you do not have eight Hopper GPUs, since the card lists that as a requirement.

Where to use it:

• A server with eight NVIDIA Hopper GPUs
• Python 3.12 with a recent CUDA toolkit
• A Docker container from the published image

Use case | Who | What they get back
Text to video | Researchers | A 10-second clip from a written prompt
Image plus text to video | Researchers | A 10-second clip that starts from your still image
Sound with the picture | Anyone generating clips | An audio track added into the output file

How to set it up: Check the hardware first, then choose Docker or a manual install. Using the ready-made Docker image saves you from building the compiled dependencies yourself.

• Eight NVIDIA Hopper GPUs
• Python 3.12 together with a current CUDA toolkit
• ffmpeg on the PATH, or the video is written without sound

1. Pull the published Docker image, or install the dependencies yourself.
2. Make sure ffmpeg is available if you want sound in the file.
3. Download the weights from Hugging Face.

docker pull sandai/magi-2-preview:latest

pip install -U huggingface_hub
hf download sand-ai/MAGI-2-preview

How to use it: You give the code a text prompt, or a prompt plus a still image, and it writes one video file. Generation runs in two stages. The part of the card used here does not show the run command, so read its usage section next.

Tool | What it does
magi2_preview | Denoises the clip at low resolution
magi2_refiner | Takes that result up to 1080p

Example: You supply a text prompt such as "a paper boat drifting down a rainy street", optionally with a still image to start from. What comes back is a video file ten seconds long, with generated sound added when ffmpeg is present.

For newcomers and experts: If you are new, know that this model needs a large GPU server, so it is not something to run on a laptop. A good first step is to read the requirements and confirm you have eight Hopper GPUs before downloading anything.

For experts, the model is a sparse mixture of experts, so it holds 114B parameters but activates 6B per token. A low-resolution model and a separate refiner share the work. The only clip length it supports is 10 seconds, which limits what you can generate.

Text prompt, with an optional image
magi2_preview at low resolution
magi2_refiner up to 1080p
Video with the audio track added

Good to know:

Title: Parameters in billions
Total: 114 B
Active per token: 6 B

• Hardware: eight NVIDIA Hopper GPUs, plus Python 3.12 and CUDA tools
• Clip length: 10 seconds is the only duration it supports right now
• Sound: without ffmpeg on the PATH the video is still written, just silently
• License: Apache 2.0
• Cost to run: no API key or paid service named, but it needs a large GPU server
• Data it reaches: downloads the weights and, optionally, the Docker image
• Not for: other clip lengths, or machines without eight Hopper GPUs

On Hugging Face, 13 November 2026

• Task: image text to video
• Formats: Safetensors
• Downloads in the last 30 days: 0
• Likes: 125
• License: apache-2.0
• Fine-tunes: 2
• Tags: safetensors, text-to-video, image-to-video, image-text-to-video, text-to-audio-video, image-to-audio-video, image-text-to-audio-video, audio-video-generation, multimodal, synchronized-audio-video, mixture-of-experts, magi-2-preview

Links

• Model: https://huggingface.co/sand-ai/MAGI-2-preview
• Files: https://huggingface.co/sand-ai/MAGI-2-preview/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download sand-ai/MAGI-2-preview

Then follow the usage in its model card. Nothing in this kit runs by itself.