Chinese startup Qingyan raises $15M Series A to build math-first world models
What's the deal? Qingyan Technology, a four-year-old Chinese startup founded by mathematicians, has raised a 100 million CNY (≈$15 million) Series A. Legend Capital led the round, with Ningbo Yishi and Suzhou Suchuangtou participating.
What's the endgame? The company wants to build a "world model" — AI that can understand and predict how the physical world evolves. Its pitch is to start from fundamental mathematics, using differential geometry to describe objects and verify whether their properties and motion obey physical laws.
Why now? World models are among 2026's most contested concepts, with video, VLA, and self-driving companies all claiming the label. Qingyan's team argues no one has pinned down a rigorous definition; it thinks the gap is a lack of structured, verifiable methods for representing physics — "exactly where mathematics can intervene."
The startup was incubated by Tsinghua University and the Beijing Yanqi Lake Institute of Applied Mathematics. Its name combines characters from both. Chairman Sun Mingming and board member Tang Ke are researchers there; Zhang Yingwei is chief executive officer and Wang Fan is chief technology officer.
What's the plan? Qingyan builds around two pillars: a "geometric-physical representation and verification" system, and a data flywheel it calls a "model-data resonance loop." Because physical AI data is scarce and costly to gather, it is building two self-run collection sites in Beijing this year, developing its own capture hardware, data pipelines, and labelling systems for 4D and 5D data.
The company has already reached some commercialisation. At an embodied-AI training ground in Qinhuangdao, led by the government and hosting several robotics firms, Qingyan acts as a neutral third party running the underlying data infrastructure. It says its self-built physics engine is 5.2 times more computationally efficient than mainstream engine Newton at comparable precision.
What could go wrong? The technical path is far from settled. Wang argues world models will diverge over how much of a scene to reconstruct, and Sun notes some estimates put the data needed to match large-model performance at trillions of samples — a scale he calls unrealistic under current paradigms.
The signal: Legend Capital frames Qingyan as a candidate to be China's paradigm-defining "Neo Lab," echoing Silicon Valley research shops like OpenAI and SSI. Tang sees a generational shift: "the last generation of internet startups was product managers starting companies; this one is people who understand technology." As mathematicians increasingly join model firms, investors are betting the next 0-to-1 breakthrough will come from deeper, more basic math.
Read more: 36kr.com
Image credit: Generated with Gemini
