Summary
MAGI-2 Preview is an intermediate research release from Sand.ai that explores scalable unified audio-video generation. It studies how architecture, training systems, and multimodal data can work together at the 100B parameter scale.
MAGI-2 Preview is an intermediate research release from Sand.ai that explores scalable unified audio-video generation. It studies how architecture, training systems, and multimodal data can work together at the 100B parameter scale.
The model uses a single-stream design with a 40-layer backbone, a sparse middle made of multi-head mixture-of-experts layers, and a smaller activated expert set. The training system focuses on communication, kernels, routing stability, optimization, and memory efficiency.
MAGI-2 is intended for researchers investigating large-scale generative video, physical plausibility, long-term consistency, and audio-video synchronization. The preview is explicitly a research milestone, so its results should be treated as evidence about scaling direction rather than a finished commercial product.
Get more likes & reach the top of search results by adding this button on your site!


Get top updates in AI to your inbox every weekend. It's free!