--- title: "Zhu Jun's Vidu: the Tsinghua lab that chased Sora and passed it" date: 2026-10-08 category: Builders site: NeuroAI canonical: https://neuroai.site/a/na-people-zhu-jun-shengshu-vidu language: en --- # Zhu Jun's Vidu: the Tsinghua lab that chased Sora and passed it > How a Tsinghua professor turned a university research group into Shengshu, the studio behind China's Vidu video model. In January 2026 an AI video model built by a Beijing startup topped an international leaderboard, landing second in the world behind only xAI's Grok. Two years earlier, almost no one outside the city had heard of its maker, Shengshu (生数科技). Even fewer knew the Tsinghua professor behind it: Zhu Jun (朱军). ## The researcher who would not wait for industry Zhu was born in January 1983 in Funan County, Anhui province. He earned bachelor's, master's and doctoral degrees from Tsinghua University, finishing his PhD in computer science in 2009, then spent two years as a postdoctoral researcher at Carnegie Mellon University. Back at Tsinghua he became a full professor, was named an IEEE Fellow in 2023 and an AAAI Fellow in 2024, and serves as vice dean of the university's Institute for Artificial Intelligence. His doctoral adviser was Zhang Bo (张钹), a member of the Chinese Academy of Sciences. Zhu's group, TSAIL, spent years on probabilistic generative models — the math underneath modern diffusion systems. A 2022 paper on fast diffusion sampling was later adopted in tools such as DALL·E 2 and Stable Diffusion. By 2026 he had published more than 650 papers with over 28,000 citations and had been elected an ACM Fellow. His standing gave Shengshu unusual credibility: at its core it was his TSAIL research group spun out with operating partners, not a startup hiring from scratch. ## Building Vidu to answer Sora In March 2023, as generative-AI funding surged after ChatGPT, Zhu founded Shengshu (生数科技) and took the role of chief scientist. The bet was specific: build a *video generation model* that could match OpenAI's Sora rather than compete in an already crowded text-chat market. The core architecture, **U-ViT**, combined a Transformer with a diffusion model — and Zhu's team argued it had been in development before Sora's February 2024 reveal. Vidu was first shown at the Zhongguancun Forum in April 2024 and opened to global users in July that year. It was billed as China's first long-duration, high-consistency text-to-video model, able to render 16-second clips with coherent motion and camera moves. ## The leaderboard moment In January 2026, Vidu Q3 Pro appeared on the Artificial Analysis video leaderboard ranked **first in China and second in the world**, behind only xAI's Grok and ahead of Runway Gen-4.5, Google Veo 3.1 and OpenAI Sora 2. It was the first model to output synchronized audio and video up to 16 seconds, with on-screen text rendering and shot control. On the VBench benchmarks it had already taken double firsts in text-to-video. Commercial traction followed. Shengshu said Vidu passed US$20 million (about 1.4 billion RMB / HK$155 million) in annualized recurring revenue within eight months of launch, serving users in more than 200 countries. ## From pixels to the physical world Zhu has been explicit that video is a step, not the destination. In December 2025 Shengshu and Tsinghua released **Motus**, a "world action model" meant to give robots a unified language of vision, text and motion — effectively a brain for *embodied* AI. The company is also reported to be weighing a Hong Kong IPO that could raise more than US$500 million (around HK$3.9 billion), working with CICC and CITIC. Funding has been brisk: an angel round above 100 million RMB (roughly US$14 million / HK$109 million) in 2023 drew Ant Group and Baidu Ventures, and later rounds lifted the valuation toward 120 billion RMB (about US$1.65 billion / HK$12.8 billion). Day-to-day management now sits with CEO Luo Yihang, a former ByteDance Volcano Engine executive who joined in March 2025, while co-founder Tang Jiayu serves as president. ## A crowded field with real rivals Vidu does not compete in a vacuum. China alone fields strong video models from Kuaishou (Kling), ByteDance (Doubao) and Alibaba (Tongyi), while OpenAI and Google set the international bar. Zhu's answer is to treat video as training data for a broader "world model" — the bet that whoever models physical reality best will own the next layer of AI, not just the clip generator. That logic is what pushes Shengshu from screens toward robots. ## What readers can do now - If you produce video, run Vidu Q3 Pro against Sora 2, Runway and Kling on a real brief before assuming the Western tool wins. - For founders, note Zhu's path: a university lab turned company by keeping the research group intact and hiring operating talent, not by abandoning academia. - Track the reported Hong Kong listing — a priced deal would set a public value on China's video-generation field. ## Honest limitations This profile is based on Sixth Tone, Bloomberg (via Chinese outlets), Artificial Analysis and company statements; the leaderboard ranking is the platform's own metric, not an independent verdict on creative quality. The reported HK$3.9 billion IPO is *under consideration* per Bloomberg, not a filed or priced offering. Zhu's exact shareholding and the precise current valuation are not publicly confirmed. Dates for Vidu's global launch (July 2024) and the Q3 Pro ranking (January 2026) are consistent across sources, but benchmark scores shift monthly. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-people-zhu-jun-shengshu-vidu Free to quote with attribution and a link to the original.