Shengshu Technology

What the company is

Shengshu Technology is a Chinese generative-AI company founded in March 2023^[1]. It describes its work as two model lines: world generation models, among them its video model Vidu, and world action models, among them Motubrain^[1]. The company says its team comes from Tsinghua University, Peking University, Imperial College London and Carnegie Mellon University^[1], and it lists branch offices in Beijing, Shanghai, Hong Kong, Shenzhen, Anhui and Wuxi^[1]. Luo Yihang is chief executive and Zhu Jun is founder^[2].

The split between generation and action is a useful way to read the company. One line produces video and images; the other is aimed at models that act, which is the direction taken by much recent work on agents and robotics. Both are built on the same underlying idea of learning a model of how the visual world behaves.

Technical background

The company says it proposed the U-ViT architecture in September 2022, three months before OpenAI's DiT^[2]. Diffusion models generate images by learning to reverse a process that adds noise step by step, and the architecture that carries out that reversal largely determines how well a model scales; the company treats this work as the technical starting point of the video models it later released^[2].

Vidu

Shengshu released Vidu at the Zhongguancun Forum in April 2024, presenting it as China's first text-to-video model benchmarked against Sora^[2]. The model went online globally in July 2024^[2]. Vidu then introduced reference-to-video generation, in which supplied images, rather than text alone, anchor what appears in the output; the company says this was a world first and that it solved multi-subject consistency for commercial video, meaning that several characters or objects can be kept looking the same across a piece of footage^[2].

Consistency is the practical obstacle in video generation for advertising, animation and short drama work: a clip that looks convincing for a few seconds is of little use if the same character changes face or clothing from shot to shot. Reference-based generation attacks that problem by fixing the subjects in advance.

Adoption

By September 2025 Vidu had more than 30 million users and 6,000 developers and enterprises across over 200 countries and regions^[3]. More than 400 million videos had been generated on the platform by then, over 100 million of them through reference-to-video or reference-to-image^[3].

Funding

The company raised repeatedly over a short period. Its Series A round of several hundred million yuan, led by Bohua Capital, was announced on 19 September 2025^[3]. A Series A+ round of more than 600 million yuan, led by Zhongguancun Science City and Xinglian Capital, was announced on 5 February 2026^[2]. A Series B round of nearly 2 billion yuan, led by Alibaba Cloud, was announced on 10 April 2026^[4].

Open source

In December 2025 the company open-sourced TurboDiffusion, which it says generates a five-second video in 1.9 seconds on a single RTX 5090^[2]. It also open-sourced Motus, a unified world action model, in the same month^[4]. Releasing inference code that runs on a single consumer graphics card is a way of putting a video model into the hands of developers who cannot rent large clusters, and of setting a speed benchmark others can check.

Where it sits

Shengshu is one of a small group of companies that turned text-to-video from a research demonstration into a product with paying users. Its trajectory runs from an architecture paper, to a model shown publicly in 2024, to a global service a few months later, to a specialised commercial feature in reference-to-video, to a user base measured in tens of millions. The funding sequence — a few hundred million yuan, then over 600 million, then nearly 2 billion — tracks how capital-intensive video generation became once it moved from demos to serving traffic worldwide. The two-line framing of generation and action models suggests the company expects the same visual world knowledge to feed systems that do things, not only systems that show things. Whether that bet pays off is unresolved, but it explains why a video company advertises an action model alongside its film-making tool.