News
The world's first 100-billion-level MoE video model is open source! Sand.ai finds a new way for video generation scaling
2 min read
Source: zhidx.com
Intelligent Things Author | Wang Han Editor | Mo Ying In the first week of August, the domestic video generation track was particularly busy. On July 31, MiniMax released the first open source multi-modal generation model H3; on the same day, ByteDance launched Seedance 2.5, extending the single generation time from 15 seconds to 30 seconds. According to estimates by international consulting firm Grand View Research, the global AI video market is expected to grow rapidly from US$4.55 billion in 2025 to US$42.29 billion in 2030, with a compound annual growth rate (CAGR) of 32.2%, showing extremely high growth potential. With the continuous bombardment of new models and the imagination of a market worth hundreds of billions, the AI video generation track has ushered in an unprecedented highlight moment. During this window period, Sand.ai, a Beijing-based AI video generation startup, quietly dropped a depth bomb - the world's first 100-billion-level MoE video generation model MAGI-2 Preview was officially released and open sourced. The total parameters of this model are about 114B, and a single forward activation has about 6B parameters. On the Image to Video list of Artificial Analysis, which is most concerned about updates, MAGI-2 Preview entered the top ten in the world, ranking sixth. Excellent list results and cool use cases are the easiest things to present to the market, but the greatest value of this model is that it has proven a clear path for large-scale MoE video model expansion: a unified single stream is responsible for the joint modeling of audio and video, fine-grained MoE is responsible for expanding capacity, and the self-developed system is responsible for stable operation of hundreds of billions of parameters. The secret lies in the three-layer system of model layer, architecture layer and system layer. 1. The video generation track is hot, but the technology has hit two walls. In the first five months of 2026, the domestic AI short drama market exceeded 22 billion yuan, and is expected to hit 40 billion yuan throughout the year, with more than 600 million users. The video generation track is undergoing a critical transition from "can it be done" to "can it be done well and cheaply". demand side