A group of mathematicians set rules for world models and secured a financing of hundreds of millions of yuan.

What is the most sought-after concept in 2026? I believe the world model must be on the top of the list. Earlier this year, an entrepreneur told me that if a project does not carry the label of "world model", it cannot even get into investment review meetings, and the investment decision-making layer will not even spare a glance at it.

Looking at the industry side, my feeling is that the term "world model" is now extremely ambiguous. Video generation companies, VLA companies and autonomous driving companies all claim that they are building world models. Even though this concept is extremely popular, no one can clearly define it, which is quite like the saying that "a thousand entrepreneurs have a thousand versions of the world model".

As a result, a group of mathematicians decided to step forward. They want to find the real world model starting from the most fundamental underlying logic. Behind this group of mathematicians is a four-year-old company, Qingyan Technology. Not long ago, Qingyan Technology officially announced the completion of a Series A financing of over 100 million yuan, led by Legend Capital, with participation from Ningbo Yishi and Suzhou Venture Capital Group.

Recently, we talked to core team members of this company: SUN Mingming, Chairman of Qingyan Technology and researcher of Beijing Yanqi Lake Institute of Applied Mathematics; TANG Ke, Director of Qingyan Technology, researcher of Beijing Yanqi Lake Institute of Applied Mathematics and professor of Tsinghua University; ZHANG Yingwei, Director and CEO of Qingyan Technology; and WANG Fan, CTO of Qingyan Technology.

Starting from Mathematics

The story of Qingyan Technology starts with the two characters that make up its Chinese name "Qingyan". The company was co-incubated by Tsinghua University and Beijing Yanqi Lake Institute of Applied Mathematics, with the first character "Qing" taken from "Tsinghua" and the second character "Yan" taken from "Yanqi Lake".

For a long time after its establishment, the team kept a very low profile, focusing most of their energy on exploring data infrastructure and the integration of mathematics and AI. Gradually, the team spotted a trend: the most intensive users of data infrastructure are shifting from institutions to models. The further data capabilities move up the value chain, the more they serve models. "The combination of data infrastructure and models is getting tighter, and world models are an inevitable trend," ZHANG Yingwei recalled.

This judgment is also confirmed on the academic side. Qingyan is rooted in the Beijing Yanqi Lake Institute of Applied Mathematics, and maintains high-frequency academic exchanges with frontline mathematics researchers, which allows the team to identify the entry point for differential geometry and mathematical physics in the development of world models.

The current predicament of embodied intelligence just gives these ideas a practical foothold. Robot data is extremely scarce, and the problem of cross-ontology cannot be solved by piling up more data in theory. Some rare but fatal corner cases must be collected through targeted design. From the perspective of the Qingyan team, the root of these difficulties does not lie in insufficient data collection, but in the lack of a set of underlying methods that can structurally represent and verify the physical world, which is exactly where mathematics can play a role.

In 2026, as embodied intelligence and world models gain popularity, a truly huge problem space has emerged: enabling machines to understand and deduce the evolution of the physical world.

For the Qingyan team, what makes this timing particularly special is that if AI is to truly move into the physical world, it cannot bypass three things: "models that can understand the physical world, collection and governance of real-world data, and high-speed iteration between the two", which exactly aligns with the two core development lines of Qingyan.

On the one hand, it requires mathematical tools such as differential geometry to turn "understanding physics" from a slogan into a problem that can be modeled and computed; on the other hand, it requires real-world data supply and data infrastructure. What is more rare is that the two drive each other to form a closed loop.

Why can Qingyan do this well? On the world model track, the team believes there are two barriers that are difficult for others to replicate: one is the mathematical gene, and the other is the data flywheel.

Most world model companies on the market are engineer-driven, while Qingyan Technology hopes to use more underlying mathematical principles to change the paradigm of physical AI. Its core narrative is "starting from basic mathematics, pursuing the 0-to-1 algorithm breakthrough of physical AI".

Specifically, it uses geometric and physical methods to describe objects, and then verifies whether the properties, motion and interactions of objects conform to physical laws. This system is named "Geometric-Physical Representation and Verification System".

As explained by SUN Mingming, the geometric-physical layer "transforms objects from the geometry of pixels and points into a continuous, differentiable and computable physical state. In this state, the identity, geometric structure, pose, motion, material and various contact relationships of objects are represented in a unified manner."

He further supplemented the underlying mathematical intuition: geometry is essentially the study of invariants. No matter from what perspective or under what lighting conditions an object is observed, its geometric properties remain unchanged. With such invariants, the learning goal of the model changes from "remembering countless variations" to "grasping the constant structure", and the learning space and required data volume will be significantly reduced.

Equally important is verifiability. When the world is modeled as a geometric-physical state, the model's prediction results can be subject to rule checks and simulation verification, which provides a more specific analysis entry for error diagnosis. For physical parameters that can be identified through observation, calibration can also be carried out combined with real-world feedback.

This system is not just a theoretical concept. At the World Robot Conference in August this year, Qingyan has demonstrated a number of phased demos based on geometric-physical representation, and released a set of data: under the same accuracy, the self-developed physics engine achieves 5.2 times the computational efficiency of the mainstream physics engine Newton, and the computational efficiency of flexible body deformation is one order of magnitude higher.

Data Gene

Now let's talk about data.

It is a consensus in the industry that the biggest difference between physical AI and large language models lies in the lack of data. The precipitation of human civilization across the entire Internet provides a large amount of free corpora for large language model training, while data for physical AI has to be collected, which is expensive and scarce. Embodied intelligence companies should have a deeper understanding of this point.

Facing the scarcity of embodied data, Qingyan's solution is to build its own data collection sites. This year, two self-built collection sites will be launched in Beijing one after another, with the scale still under planning. Around the collection sites, Qingyan has self-developed collection equipment, data pipelines, automatic annotation governance and quality inspection evaluation systems to collect 4D and even 5D data (multi-modal data including time sequence, depth, tactile sense, etc.).

The benefits of self-construction can be directly reflected by figures. According to ZHANG Yingwei, CEO, a large number of self-developed innovative collection devices are applied in Qingyan's model training process, so building its own training sites can greatly improve efficiency.

There is also a type of data that becomes more expensive the scarcer it is, yet extremely critical, such as rare events like car collisions that have an extremely low probability of occurring but cause huge losses once they happen. "If robots are to enter human society, this type of long-tail data must be obtained as much as possible, and collected through targeted design."

Therefore, Qingyan has built a closed loop of "models driving data, and data feeding back to models", which Qingyan calls the "Mathematics-Data Resonance Closed Loop". The logic is that once an algorithm breakthrough is made, data collection, training, evaluation, feedback and re-collection can be carried out immediately for the new algorithm. In TANG Ke's words, no matter whether the final breakthrough is achieved by Qingyan, this highly efficient closed loop will "keep us at the table" in the competition.

The base of this collection and data platform is built on top of data spaces and data sandboxes, which fully meets the national requirements for data compliance and trusted delivery. This is exactly the institutional accumulation that startups are difficult to make up for in the short term.

Around this data closed loop, Qingyan has built three business curves: the first is data infrastructure, which exports data spaces and data asset management platforms to external parties; the second is data collection sites. In addition to self-construction, Qingyan also exports collection equipment and platforms to external parties, and even undertakes the transformation of existing data collection centers; the third curve is models, including world models (understanding and deducing world evolution), action models (planning, decision-making, control and ontology adaptation), and synthetic data based on the geometric-physical engine.

This accumulation also allows Qingyan to achieve commercialization relatively quickly. In Qinhuangdao, in the embodied intelligence data training site led by the government and settled by multiple embodied intelligence companies, Qingyan plays the role of an independent third party, responsible for the underlying data infrastructure.

Robot data from different manufacturers needs to be managed and transmitted separately, and asset values need to be accounted for. This role can only be undertaken by a neutral third party that does not touch application scenarios. Similar demands are emerging spontaneously. "Many completed data collection centers in various regions are now communicating with us to use our platform for transformation due to operational problems," said ZHANG Yingwei.

However, Qingyan's commercialization scenarios are not only focused on embodied intelligence. SUN Mingming mentioned, "The current predicament of embodied intelligence is actually an external manifestation of the gap in environmental perception and decision-making capabilities of the entire AGI system. The foundational models in the CV field are far from reaching the capabilities of large language models. After this gap is filled, the digital world, portable AI devices, and even AI for science will produce products completely different from those we see today."

A detail that is easily overlooked is that the understanding of cognition in this track is diverging. WANG Fan gave an analogy: when humans think, they will automatically focus on relevant objects and ignore the surrounding environment, while many current world models try to reproduce every detail in the field of vision. "This line of development will definitely diverge. For robots, the key is not to reconstruct the entire scene in every detail, but to perceive the world and make decisions based on that." In his view, the technical routes of world models are far from converging, and new possibilities at the architecture level are gradually reaching a consensus.

Chinese-style Neo Lab

Although Qingyan has achieved a certain degree of commercialization, what makes me most curious about this team is its Neo Lab gene.

We have heard a lot about the Neo Lab stories in Silicon Valley. Whether it is the earliest OpenAI, SSI founded by Ilya Sutskever, Discovery Loop founded by Jeff Dean, or AMI Labs founded by Yann LeCun, all of them have secured hundreds of millions of dollars in financing with a vision and a group of outstanding talents.

China seems relatively conservative and cautious in this regard, so we wrote in the previous article *An Experiment Worth 140 Billion Yuan* that p7k founded by LIN Junyang this year may be China's first independent, VC-supported, paradigm-pioneering AI Lab.

The view that defines Qingyan as a "Neo Lab" is exactly put forward by Legend Capital, the lead investor of its Series A round. In their view, Qingyan has the opportunity to become China's paradigm-pioneering AI Lab. Compared with the above-mentioned Silicon Valley cases, Qingyan is derived from universities and research institutes, focuses on the cutting-edge of technology, engages in practical and specific businesses, and promotes the industrialization of scientific research capabilities of universities.

SUN Mingming broke down two prerequisites for the establishment of a Neo Lab.

First, the lab must face up to a cutting-edge public problem. The current data in the embodied intelligence / Physical AI field is extremely scarce, and the tasks are extremely difficult. "To achieve the effect of large language models, some people estimate that trillions of levels of data are needed, which is obviously unrealistic. The paradigm of large language models cannot be successfully replicated by simply copying. What paradigm can lead this field to real success? This is the practical problem that plagues the whole industry, and also the original intention of our establishment."

Second, it must be a new organizational paradigm. "Why can't this be done in large companies, research institutes or universities? Essentially, Neo Lab carries a new organizational form in the AI era: flat management, open-ended problems, and bottom-up exploration."

TANG Ke regards this as a replacement of the previous generation of entrepreneurship paradigm. "The previous generation of Internet entrepreneurship was led by product managers who are good at product design. This generation of entrepreneurship is led by people who master core technologies. Knowledge is becoming a scarce product and a competitive barrier, and the source of knowledge lies in research institutes and universities. A large number of companies of this new generation will grow out of research institutes and universities."

TANG Ke shared an observation of his: investors are paying more and more attention to underlying breakthroughs in basic disciplines. World model research has strong constraints, for example, data is hard to obtain, and the problems to be solved are sufficiently complex. "The more this situation is, the more the industry hopes to find answers from the underlying, more fundamental mathematics."

As supporting evidence, in the past two years, more and more mathematicians have joined large model companies and achieved success. The current market is waiting for a new 0-to-1 solution in terms of technical paradigm.

Similarly, this label has also become a scarce asset in the eyes of investors. It would be best if this can represent a breakthrough made by China. Investors believe that just like large language models, there will definitely be China's own unique world model paradigm and corresponding leading companies.

Another plus point lies in commercialization: Qingyan has obtained real commercial revenue while conducting basic research, so it does not have to follow the path of those overseas research institutions that raise financing simply by gathering a group of cutting-edge scholars.

When asked about the competitive variable he is most worried about, WANG Fan answered very directly: "What I am most worried about is that we fail to make sufficient breakthroughs before the technical routes converge." He further explained that the current embodied intelligence and world model field is in a stage of diverse development. Once the path converges, the competition will become a relatively boring resource game, which is not a favorable battlefield for a company that has not achieved a cliff-like leading advantage.

SUN Mingming extended the perspective to a longer time scale: "The final solution to this problem will definitely be explored jointly by several Neo Lab-like companies, just like large language models are developed through joint exploration by several companies. OpenAI pioneered the path of large models, o1 and chain of thought, and Anthropic contributed the paradigm of agentic and coding capabilities. We hope to become one of the companies in this exploration pattern."

Whether these visions can be fulfilled remains to be tested by time. According to the plan disclosed by SUN Mingming, a publicly accessible model will be released by this time next year. Whether the geometric-physical engine can stand firm in real tasks will be the first visible answer.

This article is from WeChat Official Account "China Investment Network", written by ZHANG Xue, and published with authorization from 36Kr.