Video generation startup Sand.ai (founded January 2024) has raised over $100 million across two funding rounds. Investors include Look Capital, Lollapalooza Capital (Wang Huiwen's family office), Jiukun Ventures, Matrix Partners China, MSA Capital, Innovation Works, Source Code Capital, IDG Capital, and Baidu Ventures. Xinghan Capital served as financial advisor.
Sand.ai founder Cao Yue told the press that the team has stuck with what many see as an unconventional approach: autoregressive video generation, rather than the dominant diffusion-based methods. The company's earlier Magi-1 model topped Google DeepMind's Physics-IQ benchmark for physical plausibility.
To break what the team calls the "impossible triangle" of video generation — cost, speed, and quality — Sand.ai began exploring Mixture of Experts (MoE) architectures last year. It plans to release its next-generation video generation model in July 2026 (Q3), featuring MoE for efficient inference and the largest parameter count among open-source models. That model will be open-sourced.
On the commercialization side, Sand.ai is pursuing a dual-pronged strategy of model and product development. Its music agent product VidMuse, launched in January this year, hit $10 million in ARR within just two months. Meanwhile, its open-source MagiAttention operator library is now used by almost all multimodal model teams in China and has received an official recommendation from NVIDIA.
On the hotly debated concept of "world models," Cao Yue believes we're still in the pre-GPT era (before GPT-1), with both data and methods far from converged. He argues that video is the most important data modality for reaching a world model — and that the right path is to let models learn physics by predicting raw pixel/frame observations, rather than imposing human priors to explicitly model state variables.