MetaCircle Secures Hundreds of Millions in Funding: Using a "Meta-Model" to Give AI "Intuition"

In 1998, an assistant professor from Tsinghua University's Institute for AI founded a startup and secured two funding rounds in four months.

Author: Dong Zibo Editor: Zhang Xian

Leiphone.com · Jingxi has learned that MetaCircle, established only four months ago, has announced the completion of two consecutive funding rounds totaling several hundred million yuan. The funds will be used for core algorithm iteration of the meta-model and exploration of interim products, team expansion, and continued release of self-developed benchmarks. Among them, the company's seed round was led by Shunwei Capital, with participation from Shanghai Future Industry Fund, Xinglian Capital, Junli Capital and other institutions; the seed+ round attracted industrial parties and state-owned background institutions such as Shanghai Semiconductor Industry Investment, 360 Strategic Investment, Shoucheng Capital, and the Shanghai AI Industry Series Fund, as well as well-known dual-currency funds such as BV Baidu Ventures, Shanzing Capital, and Yunshi Capital. Today the whole world uses large language models (LLMs) to "think," yet problems such as scientific discovery, modeling the physical world, and the design of AI itself are difficult to solve through language alone—what models need is a more fundamental intuition. The recent release of models such as Jev has also allowed the capital market to see for the first time the feasibility of the "intuition model" route. What MetaCircle is building, the "Meta Model," is an "intuition machine" that can directly generate model architectures and training hyperparameters: give it an unfamiliar model, and it can predict how it will perform and in what scenarios it is better suited—equivalent to installing a "god's-eye view" for the AI industry, which is precisely the fundamental difference between MetaCircle and the popular RSI (Recursive Self-Improvement) route. In the view of Liu Ziming, founder and chief scientist of MetaCircle, using large language models for AI self-evolution is like exploring an off-trail route outdoors: no map, no directional intuitive basis, and the only way to clear it is to burn massive compute through repeated trial and error; whereas the meta-model can learn to directly generate the underlying patterns behind the performance of AI models, equivalent to obtaining the entire map and being equipped with GPS, with first-order and even global informational guidance. Liu Ziming was born in 1998 and is currently an assistant professor at Tsinghua University's Institute for AI; he studied undergraduate at Peking University's Department of Physics, received his PhD from the Massachusetts Institute of Technology (MIT), and previously conducted postdoctoral research in neuroscience and AI at Stanford University. His most important prior achievement, KAN (Kolmogorov-Arnold Networks), is also the technical core of today's "meta-model": it replaces the fixed activation functions of traditional neural networks with "learnable activation functions," turning models from a "black box" into a "white box"—since its release in 2024, it has gained more than 16,000 stars on GitHub and more than 5,700 Google Scholar citations. Next year, the company will launch a general meta-model 1.0, not limited to scenarios, to generate specialized models adapted to various vertical domains. The team's self-developed benchmark for testing model-training intuition, "Architecture IQ," will also be released soon, to evaluate the level of intuition of large models and human researchers regarding model training.