MAGI-2 Preview (0912)

MAGI-2 Preview (0912) is a unified audio-video generation model developed by Sand.ai that explores an efficient path to scaling video generation architecture. Built on a sparse mixture-of-experts (MoE) framework known as MagiMoE, the model features a total parameter count of 114B while activating approximately 6B parameters per token during inference. It is designed to process text prompts or text combined with still images to generate cohesive video clips with synchronized audio tracks.

The architecture relies on a single-stream design where text, video, and audio tokens are concatenated into a unified sequence processed directly by the Transformer backbone. This integration allows modalities like language, facial expressions, body motion, and sound effects to exchange information continuously throughout inference rather than depending solely on restricted cross-attention modules. Generation proceeds in a two-stage pipeline, where a preview stage denoises the clip at a lower resolution before a refiner stage upscales the output to 1080p resolution.

MAGI-2 Preview natively handles synchronized audio-visual generation and complex camera movements, supporting 10-second clip durations as its standard output format. The model captures shot scale, composition, perspective, and motion intents such as zooming and tracking to closely mimic real-world cinematography. Distributed through open weights, it provides a foundation for researching large-scale efficient generative video models.