MetaCircle Raises Back-to-Back Rounds to Build AI That Designs AI, Led by KAN Author Liu Ziming

A four-month-old Chinese startup is betting that the next layer of the AI stack is not a bigger model but a system that designs models. MetaCircle, known in Chinese as Yuanhuan Zhineng, said it has closed two consecutive funding rounds totaling several hundred million yuan since its founding, led in the seed round by Shunwei Capital with the Shanghai Future Industry Fund, Xinglian Capital and Junli Capital participating.

The seed-plus round pulled in a notably industrial bench: Shanghai semiconductor industry investment vehicles, a 360 strategic investment, Shoucheng Capital and Shanghai AI industry funds, alongside BV Baidu Ventures, Shanxing Capital and Yunshi Capital. Qiguan Capital is serving as long-term financial advisor. The company did not disclose exact amounts or valuations.

The founder is the draw. Liu Ziming, born in 1998, is an assistant professor at Tsinghua University's Institute for AI with a physics undergraduate degree from Peking University, a PhD from MIT and a Stanford postdoc in neuroscience and AI. He is best known as creator of KAN, the Kolmogorov-Arnold Networks architecture that replaced fixed activation functions with learnable ones and has accumulated more than 16,000 GitHub stars and over 5,700 citations since 2024.

His thesis is deliberately contrarian to the scaling narrative. Large language models, he argues, cover only about 30% of the world's problems — text, math and code, the tasks that can be put into words. The remaining 70%, from scientific discovery to physical-world modeling to the design of AI itself, needs something lower-level: intuition about what architecture and hyperparameters will work before the training run is paid for. MetaCircle's meta model is positioned as that intuition engine — shown an unfamiliar model, it learns the regularities behind its performance, predicts how it will behave and in which scenarios it fits, and generates architectures and training configurations directly.

The company is explicit that this is a different route from recursive self-improvement, the RSI race that has captured the funding imagination elsewhere: rather than an AI improving itself in the loop, MetaCircle wants a system that gives developers a map of model-design space before they spend compute. Its first commercial targets are deliberately narrow — automatically designing predictive models for time-series data in industrial inspection, autonomous driving, weather forecasting and quantitative finance, domains where continuous data records are dense and domain models still get built by hand.

MetaCircle also plans to publish its own benchmark, Architecture IQ, to score how much training intuition large models and human researchers actually have. That is a convenient test for a company selling intuition, and the obvious caveat applies: the entire framing rests on the founder's research pedigree and investor conviction rather than a shipped product, and none of the capability claims have been independently verified. But the round structure — state-backed semiconductor and AI funds alongside Shunwei and Baidu Ventures — suggests Chinese capital is willing to fund the AI-for-AI thesis at the earliest possible stage, following the path Jev and the decision-model wave cleared for intuition-first systems.