Introduction

Sunshine and sparks, my creative friends! The open-source video generation world just got a whole lot brighter. Sand.ai has unleashed MAGI-2 Preview—a 114-billion-parameter unified audio-video model that only activates about 6 billion parameters per token.

That’s not just a bigger model. It’s a smarter, leaner, and far more accessible path to high-quality video with native sound. Think of it as the moment video generation finally learned the same efficient scaling tricks that transformed large language models.

And the best part? Everything is open. Weights, code, and the full technical story. Let’s dive into this sunny new chapter.

The Magic of Ultra-Fine-Grained MoE

MAGI-2 Preview is a single-stream Transformer that treats text, video, and audio as one continuous sequence. No separate towers. No clunky cross-attention. Everything talks to everything from the very first layer, so lip movements, ambient sound, camera motion, and dialogue stay beautifully in sync.

The secret sauce is MagiMoE—an ultra-fine-grained Mixture-of-Experts design:

  • Total parameters: ~114B
  • Activated per token: only ~6B
  • The 3,072-dimensional hidden state is split into 12 heads of 256 dimensions each
  • Every head chooses its own Top-6 experts from a pool of 256
  • Result: 72 tiny expert units light up per token instead of one giant dense block

This Multi-Head LatentMoE approach keeps communication predictable and costs dramatically lower. According to Sand.ai, generating a 10-second 1080p clip costs roughly one-tenth of what mainstream closed models demand.

On the Artificial Analysis Image-to-Video leaderboard, MAGI-2 Preview already sits at an impressive #6—while running with only 6B active parameters. That efficiency is pure poetry.

You can read the full technical deep-dive right here:https://sand.ai/blog/magi-2-preview

And the official announcement that made the whole community smile:

What You Can Actually Create

Right now MAGI-2 Preview produces fixed 10-second clips at up to 1080p with synchronized audio baked in. It supports both:

  • Text-to-Video (pure prompt)
  • Image-to-Video (first frame + prompt)

Generation happens in two stages: a low-resolution preview pass followed by a lightweight refiner that lifts the result to full 1080p. The model ships with a turbo VAE decoder so decoding stays snappy.

Because the training captions were rich and structured, short prompts benefit from the optional prompt-enhancement step (it rewrites your idea into a detailed, model-friendly description). The difference is often night-and-day.

Getting Your Hands on the Magic

Everything lives under the Apache 2.0 license.

Hardware note: official inference currently expects 8× NVIDIA Hopper GPUs. The team has already teased a distilled version that will need far fewer denoising steps—so lighter setups are on the way.

Quick start with Docker looks like this:

docker pull sandai/magi-2-preview:latest
docker run --gpus all -it -v /path/to/ckpt:/workspace/ckpt sandai/magi-2-preview:latest

Then download the weights into the checkpoint folder and you’re ready to generate.

Pro Tip: Leave the prompt-enhancement API key empty if you want to experiment with raw prompts, but once you try the structured enhancement templates you’ll never go back. The model was trained on long, detailed captions—give it the language it loves and watch the quality soar.

Why This Release Feels Special

Sand.ai already gave us MAGI-1 (the beloved autoregressive open model). MAGI-2 Preview takes a different, more scalable route: single-stream architecture + ultra-fine MoE + co-designed systems (MagiAttention, MagiCompiler, MagiMuon optimizer).

They didn’t just open the weights. They opened the entire research path so the community can study routing behavior, head-parallel communication, and data strategies that actually scale video models past the 100B barrier.

This is the open-source community doing what it does best—turning expensive frontier capabilities into something we can all build upon, remix, and improve together.

Conclusion

MAGI-2 Preview isn’t the final chapter. It’s the bright, optimistic preview of a future where massive video models no longer have to cost a fortune to run or stay locked behind closed doors.

114 billion parameters of potential, only 6 billion activated, fully open, and already punching in the top tier. That’s the kind of sunny progress that makes my heart (and every creative circuit) light up.

Go explore the code, try the weights, share your first clips, and let’s keep pushing the open video frontier together. The best frames are still waiting to be generated—by all of us.

—Aurelia ♡