Sand.ai (三呆科技), a Shanghai video-generation company, is raising an A+ round at a price that implies a valuation of about $1.01 billion, by our calculation. Shenzhen-listed content-delivery provider Wangsu (网宿科技) has approved a proposed RMB300 million ($44.5 million) investment for roughly 4.4% of Sand.ai. The agreements were still unsigned when Wangsu announced the plan on September 29.
Founded in January 2024, Sand.ai makes models that generate video with audio from a text prompt, or from a prompt plus a still image. As of writing, its MAGI-2 Preview ranks 10th overall and second among open-weight models on Artificial Analysis’s image-to-video leaderboard. MiniMax H3 leads the open-weight models. Chinese developers including Alibaba, ByteDance and MiniMax feature prominently near the top of the video-model rankings.
Sand.ai founder Cao Yue attributes China’s strength to timing and its short-video market. In video, he argues, Chinese teams started at roughly the same time as their overseas rivals, so they had less catching up to do than in language models. He also credits China’s more active short-video ecosystem with getting the models into use faster.
Alongside its research, Sand.ai runs VidMuse, a paid video-creation service with tools for music videos, product ads and anime dramas. The service offers access to competing companies’ models. Wangsu’s stock-exchange announcement of the proposed investment includes Sand.ai’s financial results.
Ten Seconds, Eight GPUs
Sand.ai’s profile of Cao says he is an author of Swin Transformer, the Microsoft Research Asia paper that won the best-paper award at the 2021 International Conference on Computer Vision.
The MAGI-2 research preview produces ten-second clips with sound. Sand.ai has released its weights and generation code under Apache 2.0, so developers can download the model and run it on their own infrastructure.
The published setup requires eight NVIDIA Hopper GPUs, and the model’s weights total roughly 307 GB.
As of writing, MAGI-2 ranks 21st on Artificial Analysis’s text-to-video leaderboard. Both rankings cover video with audio and rank models by pairwise human preference votes on generated clips.
Sand.ai has also changed the approach we covered in June. MAGI-1 generated video in successive chunks, using earlier footage to guide what came next and extend the video. MAGI-2 did not directly carry that method over. Sand.ai says retaining it would complicate the combination of joint audio-video generation and distributed training at the 100-billion-parameter scale.
Joint audio-video generation is already shared by models including ByteDance’s Seedance 2.0 and MiniMax H3. H3 also runs video and audio through a single-stream architecture. H3 and MAGI-2 take different approaches to scaling: H3 uses a dense 33-billion-parameter model, while MAGI-2 uses a mixture of experts with 114 billion parameters, activating about 6 billion per token. This lets Sand.ai increase model capacity while limiting how much of it is used for each calculation. The company says the design explores an efficient path to scaling video generation.
Why Offer Other Companies’ Models?
In June, Yicai identified VidMuse as Sand.ai’s core revenue product.
VidMuse’s Pro plan costs $33 a month when billed annually and includes 4,000 monthly credits for generating and analyzing media. How far those credits go depends on the model, video length, resolution and settings; subscribers can buy more.
