Sand AI is a Beijing-based generative artificial intelligence company founded in 2024 and focused on AI video generation. Its mission centers on advancing autoregressive approaches to video, a technically distinct path from the diffusion-based pipelines that dominate much of the field. The company's flagship model, MAGI, is positioned as the first high-quality autoregressive video generation model, generating video frame sequences in a way designed for scalability and longer, more coherent outputs.
In April 2025, Sand AI released MAGI-1 as a fully open-source model, a notable move in a competitive video-generation landscape where most leading systems are closed. MAGI-1 emphasizes infinite scalability and high fidelity, and the company has shipped surrounding infrastructure such as MagiCompiler and MagiAttention to make large-scale autoregressive video generation more efficient. By open-sourcing the model, Sand AI sought to build developer adoption and establish credibility against larger, better-funded rivals.
Sand AI provides both an API platform and a consumer-facing product for creating videos with its MAGI models, targeting creators, developers, and businesses that want programmatic access to generative video. Its later MAGI-1.1 release continued to push output quality. The team carries research pedigree connected to the Swin Transformer line of work, lending technical credibility to its architectural bets.
The company raised roughly $50 million across early financing, with a Series A round in 2025 led by Matrix Partners China and participation from investors including IDG Capital. Like other Chinese AI video startups, Sand AI operates amid intense domestic competition and content-moderation scrutiny, but its open-source, autoregressive strategy gives it a differentiated identity in a crowded market racing to make high-quality AI video broadly accessible.