Sand.ai has open-sourced MAGI-2 Preview, a hundred-billion-parameter Mixture of Experts (MoE) video generation model, marking a potential shift in the industry's scaling approach. The model has a total of about 114B parameters but activates only about 6B parameters in a single forward computation, addressing rising training and inference costs as video models grow. It is among the few globally to apply hundred-billion-parameter MoE to a unified audio and video generation model, with its weights, code, and training system fully open-sourced. MAGI-2 Preview ranks 6th globally on the Image to Video track of Artificial Analysis. The release comes as the industry, including models from OpenAI, Google, Kling, and MiniMax Hailuo, pushes boundaries in image quality, duration, and controllability, while facing challenges from processing complex characters, actions, shots, and unified audio-video content. The previous path of scaling Dense Transformer models is approaching its limit, and MoE, which decouples model capacity from single computation volume, is proposed as a solution to reopen the Scaling Law for video generation.