The Tsinghua-affiliated startup has raised $300 million in 10 months, betting that open-source technology can help an independent startup compete with industry giants.
By Jiang Siyuan
Sand.ai has completed a $200 million Series A+ funding round, its third financing this year, bringing its total fundraising to $300 million over 10 months and making it a new unicorn in video generation.
The round was co-led by Cornerstone Capital and Wangsu Science & Technology (300017.SZ), with participation from Shanghai state-backed investors, Times Capital and existing shareholders including Matrix Partners China.
Wangsu, a leading Chinese content delivery network (CDN) and edge computing provider, has supported high-traffic services for ByteDance, Tencent (0700.HK) and JD.com (9618.HK) (JD.US). It also provided underlying network and edge technology for the Doubao AI model’s performance at China’s 2026 Spring Festival Gala.
Founded by Cao Yue, a former principal researcher at Microsoft Research Asia and a main author of the Swin Transformer computer vision model, Sand.ai is among a small number of Chinese AI startups to achieve unicorn status with video generation as the core business.
While competitors such as Aishi Technology, Shengshu Technology (Vidu) and Vivix are exploring world models and embodied AI, Sand.ai has concentrated on foundational models. It released and open-sourced its MAGI-2 Preview model in mid-2026. A larger, full version, which uses a mixture-of-experts (MoE) AI architecture, is expected to launch soon.
The capital race for video model leadership
Video generation has become one of the most commercially promising businesses in AI.
ByteDance’s Seedance 2.0 generated more than 1 billion yuan ($150 million) in monthly revenue by mid-2026, with a gross margin of around 70%. Kuaishou Technology’s (1024.HK) Kling AI reported second-quarter revenue of more than 850 million yuan, up over 200% year-on-year.
Kuaishou announced plans in July to raise independent financing for Kling, valuing the business at approximately $15 billion before the round and an estimated $18 billion afterward. The capital increase was completed in late October, with approximately $2.9 billion raised.
These results demonstrate that video models have moved beyond technical promise to generate substantial revenue. As AI video expands into advertising, short-form dramas, gaming and film production, large-scale customers are increasingly seeking multiple model suppliers to reduce costs and supply risks.
Kling’s fundraising also provides a valuation benchmark, showing investors are willing to value video AI businesses independently when they possess leading technology and commercial traction.
Cornerstone Capital investors said they believe the sector remains relatively early-stage, with significant room to improve inference speed, costs and consistency in long-form generation. As foundational models advance, new products and paid applications are likely to emerge, leaving room for multiple successful companies.
From MAGI-1 to large-scale MoE models
The long-term value of a foundational model company depends on its ability to continually develop, train, and improve models.
In May 2025, Sand.ai launched MAGI-1, an autoregressive video generation model, and open-sourced its technical framework. At the time, the startup had only 30 to 40 employees across data, infrastructure, algorithms, and evaluation. Cao later acknowledged that the team had underestimated the difficulty of the approach.
MAGI-1 generated video in sequential segments, using previously generated footage to produce subsequent clips. Each segment was created through diffusion-based denoising. The approach enabled interaction during generation but required the team to tackle challenges involving continuity between segments and the influence of earlier frames on later ones.
The startup also built its own training systems and data-processing pipelines. One result was MagiAttention, an open-source computing library that other teams can reuse. Tencent’s Hunyuan Hy-Nomas project incorporated it as an official dependency, while Nvidia’s NeMo AutoModel framework added support.
MAGI-1 attracted investor attention after achieving leading results on the Physics-IQ benchmark for video continuation, demonstrating an ability to predict physical processes. It continues to rank first on the benchmark. Sand.ai’s technical report also said an enhanced decoder version achieved the highest overall score in VBench’s image-to-video evaluation.
In August 2026, the company released MAGI-2 Preview, advancing its research into large-scale model training.
The model has 114 billion total parameters, with about 6 billion activated per token. Its ultra-fine-grained MoE architecture distributes computation among specialized components, activating only a subset during generation. This expands the model’s capacity while controlling computational requirements.
Sand.ai moved away from MAGI-1’s segmented autoregressive design, adopting a single-stream architecture that processes text, images and audio through one backbone network. The company also developed supporting systems for expert routing, parallel training and optimization, using extensive video data for pretraining before refining the model with higher-quality data and human annotations.
MAGI-2 Preview demonstrated that the architecture and training infrastructure could operate together at scale. Sand.ai is now developing the full MAGI-2, seeking to translate its accumulated research and engineering capabilities into improved model performance.
Open source offers an alternative competitive path
In the history of foundational software, open and closed have always coexisted as competing approaches. For independent model developers, open source provides a way to compete without relying entirely on proprietary products or distribution channels.
Large technology companies can promote models through established platforms and customer relationships. Startups often lack those advantages. By releasing model weights, code and technical components, they can enable developers to integrate their technology into external products and infrastructure.
The approach has gained traction in large language models. Hugging Face data showed that repositories of derivative models based on Alibaba’s Qwen series exceeded 150,000 as of July 2026. Video generation is developing a similar ecosystem: Wan2.1, for example, gained wider adoption among creators after integration with ComfyUI.
Sand.ai has released model weights, inference code and technical documentation for MAGI-1 and MAGI-2 Preview. It has also open-sourced engineering components such as MagiAttention and its Multi-Head LatentMoE design, which seeks to reduce data movement and waiting costs when activating multiple experts.
A widely adopted model could become embedded in enterprise production systems, creating revenue opportunities through deployment, hosted inference and ongoing technical services beyond the company’s own applications.
Currently, model quality, inference cost, third-party tool support and enterprise deployment capability are still changing rapidly, and the developer ecosystem has yet to settle. Sand.ai has built a foundation to compete through consecutive model releases and technical open-sourcing, with potential to contend for an influential open video foundation model position.
The startup’s three funding rounds in 10 months have bought it capital and time to compete. Its next challenge is to prove that its research and engineering capabilities can consistently produce leading models and establish lasting value for an independent company in a market dominated by technology giants.
Source:
LatePost
