Sand.ai raises $200M Series A+, reaches unicorn on MAGI video models

Alongside the co-leads, TMTPOST and Pedaily list Times Capital, Panlin Capital, Dongzheng Capital, Tianchuang Capital with Shandong railway industrial funds, Shanghai state-backed Pudong Venture Capital and Shanghai Guofang, plus follow-ons from Matrix Partners China, Huaye Tiancheng and CAS Star[1][2]. Wangsu, a listed CDN operator, had already filed that it planned to invest about 300 million yuan of its own cash for newly issued Series A+ preferred shares, framing the stake as a path into audio-video and edge inference[2].

Proceeds go to video foundation-model research, training and inference capacity, product work and overseas expansion[1]. Sand.ai plans to ship the formal, larger-parameter MAGI-2 soon after the August open-source MAGI-2 Preview[1][3].

Founded by former Microsoft Research Asia researcher Cao Yue, Sand.ai pushed autoregressive video generation and later a MoE stack instead of following the early diffusion consensus[2]. MAGI-2 Preview is described as a roughly 114B-parameter MoE with about 6B parameters active per generation; Pedaily says a distilled build can produce a 10-second 1080p clip for about 0.5 yuan, roughly a tenth of mainstream per-second pricing at the time[2]. On the product side, its VidMuse creation agent reportedly cleared $10 million ARR within two months of launch[2]. Magi remains the public face of the stack on sand.ai[4].

Video generation is still dominated by giant-owned stacks such as ByteDance's Seedance and Kuaishou's Keling. An independent Chinese lab that can raise at unicorn scale, open-source a 100B-class MoE preview, and sell a creator app gives studios and platforms a third-party API option that is not also their content competitor[2]. The formal MAGI-2 release is the next proof point for whether that bet holds.