MetaRing Intelligence has completed two consecutive rounds of financing totaling hundreds of millions of yuan, aiming to build "meta-models" capable of designing models | StarLink Portfolio Update

MetaCircle / AI × Human × Creativity

Recently, MetaCircle has completed two consecutive rounds of financing totaling hundreds of millions of yuan. The raised funds will be used for the core algorithm iteration of meta-models, exploration of phased products, team expansion, and continuous release of self-developed benchmark systems.

Among them, the seed round was led by Shunwei Capital, with participation from institutions including Shanghai Future Industry Fund, Starlink Capital, Junli Capital and other organizations; the seed+ round attracted industrial parties and state-owned background institutions such as Shanghai Semiconductor Industry Investment, 360 Strategic Investment, Shoucheng Capital, Shanghai Artificial Intelligence Industry Series Fund, as well as well-known dual-currency funds such as BV Baidu Venture Capital, Mountain Capital, and Yunshi Capital.

Right after its establishment, MetaCircle received joint support from multiple market-oriented venture capital institutions, industrial capital and state-owned funds, forming a diversified shareholder lineup.

01

What Are We Doing?

When the whole world is "thinking" with large language models (LLMs), Liu Ziming, founder and chief scientist of MetaCircle, proposed that large language models can only cover about 30% of the world's problems — clearly definable tasks such as text processing, mathematics and programming. The remaining 70% of difficult problems — scientific discovery, physical world modeling, and the design of AI itself — cannot be solved by language alone, models require more underlying intuition. The recent release of models such as Jev also allows the capital market to see for the first time the feasibility of the "intuition model" route.

The "Meta Model" developed by MetaCircle is an "intuition machine" that can directly generate model architectures and training hyperparameters: give it an unfamiliar model, and it can predict its performance and which scenarios it is more suitable for. It is equivalent to installing a "god's-eye view" for the AI industry, turning sudden flashes of inspiration into daily routines, and avoiding piling up computing power to compete for trivial advantages — helping developers avoid detours and spend computing power where it matters most.

This is the fundamental difference between MetaCircle and the popular RSI (Recursive Self-Improvement) route. In Liu Ziming's view, the industry uses large language models for AI self-evolution, which is like an outdoor wild trail adventure: there is no map, no basis for directional intuition, and the only way to make progress is to consume massive amounts of computing power for repeated trial and error.

The meta model can learn the laws behind the performance of AI models, which is equivalent to getting the entire map and being equipped with GPS, thus obtaining first-order and even global information guidance.

This judgment is supported by the founder's outstanding academic resume. Liu Ziming, born in 1998, is currently an assistant professor at the School of Artificial Intelligence, Tsinghua University. He received his bachelor's degree from the Department of Physics, Peking University, his doctorate from the Massachusetts Institute of Technology (MIT), and has engaged in postdoctoral research in neuroscience and AI at Stanford University. His most important previous achievement, KAN (Kolmogorov-Arnold Networks), replaces the fixed activation function of traditional neural networks with "learnable activation functions", turning the model from a "black box" into a "white box" — since its release in 2024, it has received more than 16,000 stars on GitHub and more than 5,700 citations on Google Scholar.

After the publication of KAN, a large number of subsequent studies around this achievement were quickly launched, which made Liu Ziming, who was in his fourth year of PhD at MIT at that time, receive a faculty offer from the School of Artificial Intelligence of Tsinghua University. "KAN is not an academic paper, nor a specific model. It is a methodology and a worldview that can generate order from chaos. It proves that model design can be converged by 'scientific research aesthetics' and first principles just like Newton's laws, and the meta model is exactly the next step of this methodology," Liu Ziming introduced.

Liu Ziming said: "If LLM is good at the 30% compressed symbolic world or linguistic world, the meta model developed by MetaCircle is to solve the remaining 70% uncompressed numerical world, whose core is the KAN model. In the short and medium term, it can be used to design dedicated prediction models for time-series data scenarios such as industrial inspection, financial quantification, meteorological prediction, power prediction, autonomous driving and world models, and verify the feasibility and commercial implementation of these models. Next year, we will launch the general Meta Model 1.0, which is not limited to specific scenarios, to generate dedicated models adapted to various vertical fields."

The company was founded in May 2026, with an average team age of about 22. Most of its members come from top universities such as Tsinghua University and Peking University, making it a typical "young genius team" that is most favored by the primary market at present. The company's slogan is "Scale Insight, not just Compute". The team will also release the self-developed test benchmark for model training intuition "Architecture IQ" in the near future, which is used to evaluate the intuition level of both large models and human researchers on model training.

02

Investor Insights

Shunwei Capital stated:

"At present, new AI paradigms are being explored all over the world. The meta model can accelerate the rapid convergence of models and spill over the capabilities of large models to other different fields. Commercially, MetaCircle first develops products with 'smaller capability boundaries': it automatically designs prediction models for time-series data scenarios such as industrial inspection, autonomous driving, meteorological prediction and financial quantification, taking the lead in verifying feasibility and realizing commercialization. Liu Ziming is a representative of young scientists of Generation Z. Whether it is AI for Science, Science of AI or AI for AI, Dr. Liu has excellent original innovation ability and unique academic aesthetic pursuit."

Shanghai Future Industry Fund stated:

"We believe that when the improvement of model intelligence becomes the main basis for model pricing, the ability to 'design the model itself' will become increasingly important. In the two routes of AI for AI, RSI is the current engineering practice, and the meta model represents a more underlying future — MetaCircle is at the forefront and early stage on this track. While most teams are pursuing to pile up computing power and polish models, they are precipitating the intuition of algorithm design, enabling AI to learn to create models. Participating in this round of investment is part of our Neolab's exploration and layout: in the earliest and most uncertain stage, we will accompany people who are truly engaged in underlying research for a long journey."

Starlink Capital stated:

"The meta model and AI for AI concept proposed by MetaCircle provide a refreshing new idea for breaking through the traditional model R&D method. We are optimistic about the long-term value of this direction, and recognize the founder team's profound accumulation and original exploration in the cross field of neuroscience and artificial intelligence. As early-stage investors, we hope to support such a team with technical ideals and the courage to challenge existing paradigms, and accompany MetaCircle to turn AI for AI from concept into reality step by step."

Yunshi Capital stated:

"The ultimate proposition of the second half of artificial intelligence is destined to form a self-closed loop of AI for AI. The success of AI for Science does not lie in piling up computing power, but in a small research team composed of a tiny number of core researchers who have top skills and extreme obsession in specific fields. Founder Liu Ziming just has such excellent scientific research taste and pure original aspiration, and is committed to developing the machine that builds models. MetaCircle deeply integrates the first principle of physics with AI for AI, and reshapes the prediction of model training laws and the full-process automated scientific research system through the new paradigm driven by meta models."