**[Sand.ai Open Sources the World's First Trillion-Parameter Video MoE Model MAGI-2 Preview]** According to monitoring by Beating from 动察, Sand.ai has released and open-sourced MAGI-2 Preview, the world's first trillion-parameter open-source MoE video generation model. The model has a total of approximately 114B parameters, but only about 6B are activated per use. It supports text-to-video and image-to-video generation, capable of simultaneously producing visuals, dialogue, music, and ambient sounds. Currently, it generates fixed 10-second content per run, with visuals upscaled to 1080p. However, larger parameters do not necessarily mean better performance. MiniMax H3, with a 33B dense model as its backbone, ranks third on the Artificial Analysis audio-visual video generation leaderboard, outperforming MAGI-2, which ranks sixth. The main highlight of MAGI-2 is its expansion of video models to trillion-parameter scale while keeping single-use activation to around 6B. The model is licensed under Apache 2.0, with full weights totaling approximately 307GB, requiring 8 NVIDIA Hopper GPUs for local operation. [Original Link]