Sand AI launches Magi 2 preview model

Sand AI has reportedly launched the Magi 2 preview model on Hugging Face, according to free.theresanaiforthat.com. Additionally, free.theresanaiforthat.com has reported that Magi has released Magi 2.

The newly released architecture is a unified audio-video generation model that features 114 billion total parameters while activating only about 6 billion parameters per token. By employing a mixture-of-experts approach known as MagiMoE, the model aims to make large-scale video generation much more efficient by decoupling total capacity from per-token computation.

Architecture and Generation Process

The Magi 2 preview model supports both text-to-video and image-to-video generation tasks, creating synchronized audio alongside the video clips. Clips are generated at a standard duration of 10 seconds. The generation pipeline executes in two distinct stages: a preview stage that denoises the clip at lower resolution, followed by a refiner stage that scales the output up to 1080p resolution.

Access and Requirements

According to the reported details, the model weights are made publicly available on Hugging Face under the repository name sand-ai/MAGI-2-preview, and the inference code is hosted on GitHub under an open-source license. Running the model locally requires significant computing hardware, specifically pointing to systems equipped with eight NVIDIA Hopper GPUs.