MAGI-2 Preview

MAGI-2 Preview is a 114B-parameter unified audio-video generation model developed by Sand.ai that activates approximately 6B parameters per token using a mixture-of-experts (MoE) design. Co-designed across architecture, systems, and data, the model explores an efficient pathway to scaling video generation by combining sparse activation with open weights. It handles both text-to-video (T2V) and image-to-video (I2V) generation tasks, outputting 10-second clips with synchronized audio tracks.

Architecture and Capabilities

Built around the MagiMoE framework, the model utilizes an ultra-fine-grained multi-head MoE layout to handle sequence lengths characteristic of video data efficiently. MAGI-2 Preview jointly generates visuals and audio, ensuring precise alignment between dialogue, lip movements, actions, and environmental sound effects. The model demonstrates advanced comprehension of cinematography, accurately reflecting camera movements such as tracking, zooming, and orbiting while managing complex spatial interactions and character performance nuances.