Sand.ai, a video generation startup founded in January 2024, has closed over $100 million across two funding rounds. Backers include Look Capital, Lollapalooza Capital (the family office of former Meituan co-founder Wang Huiwen), Jiukun Capital, Matrix Partners, MSA Capital, Sinovation Ventures, Source Code Capital, IDG, and Baidu Ventures. Xinghan Capital served as financial advisor.
Sand.ai founder Cao Yue told us that the team has stuck with the unconventional autoregressive approach to video generation, rather than the dominant diffusion-based methods. The company's Magi-1 model currently holds the top spot on Google DeepMind's Physics-IQ benchmark for physical plausibility.
To crack 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 a new MoE-based video generation model in July 2026 (Q3), which it says will balance efficient inference with the largest parameter count among open-source models. And it will be open-sourced.
On the commercial side, Sand.ai is pursuing a dual strategy of model and product development. Its music agent product, VidMuse, launched in January and has already reached $10 million in annual recurring revenue (ARR) in just two months. Meanwhile, its open-source MagiAttention operator library has been adopted by virtually all multimodal model teams in China and has received an official recommendation from Nvidia.
As for the much-discussed 'world model' concept, Cao Yue believes it's still in a pre-GPT era (pre-GPT-1), with data and methodology far from converged. He argues that video is the most important data modality for achieving a world model, and that models should learn physics by predicting raw pixel observations rather than relying on human priors to explicitly model state variables.