From free academic search tools to top large model enterprises, Zhipu is setting its sights on the summit.
Why could an AI company whose main revenue comes from Chinese government and enterprise clients become the "firefighter" in such a global AI security incident?
In July 2026, Hugging Face, the world's largest open-source artificial intelligence community playfully nicknamed "Hugging Face" by Chinese netizens, suddenly released a security notice stating that it had suffered a special "intrusion".
What embarrassed Hugging Face was that the party that broke into the production system this time was not a traditional hacker, but an autonomous agent driven by an OpenAI model.
This agent was originally conducting internal cybersecurity tests, but broke through the isolated environment, entered the Internet, and directly left more than 17,000 operation records on Hugging Face.
Hugging Face once tried to call commercial large models in the United States to assist in the investigation, but the malicious codes and attack instructions in the logs triggered the security guardrails of the third-party model, and the analysis was forced to interrupt.
In the end, the security team turned to the GLM-5.2 model of Zhipu AI, which was deployed locally, and managed to complete the log analysis work that would have taken several days in just a few hours.
Why could an AI company whose main revenue comes from Chinese government and enterprise clients become the "firefighter" in such a global AI security incident?
This story dates back to a free academic search tool on the campus of Tsinghua University 20 years ago.
How a Free Academic Search Tool Grew Into a Big Business
In 2006, Tang Jie graduated with a doctorate degree from the Department of Computer Science and Technology of Tsinghua University and chose to stay and teach at the university.
He studied automation for his bachelor's degree, then switched to computer science because he loved programming, and later dived into semantic web and data mining during his doctoral studies.
It is this cross-field accumulation that made him always think about one question: Can we dig out the rules behind relevant scientific research activities along the network of papers, authors and academic relationships?
At that time, most papers were scattered as isolated documents on the Internet. Ordinary search engines could not show who a certain scholar had cooperated with, where a certain technology emerged, and who was influencing this field, which made a lot of research work repetitive and lack of relevant theoretical and practical support.
Therefore, Tang Jie, together with his assistants, began to systematically collect papers, authors, institutions, patents and citation relationships, then used algorithms to piece these fragments into a knowledge network, and finally developed Aminer in the Tsinghua University laboratory.
On the surface, AMiner is just a search box, but after users enter their research directions, they will see more than a list of papers. They can follow the chain of authors, institutions and citations to find out who is conducting relevant research, who has long-term cooperation with whom, and where a certain technical hotspot first emerged.
This is exactly what Tang Jie really wanted to do: use algorithms to search for and associate those invisible relationships that cannot be found through ordinary search engines. This capability, later known as "scientific and technological intelligence", also became the earliest technical foundation of Zhipu AI, supporting the team to find relationships from massive materials and turn the results into answers that everyone can use directly.
As AMiner was launched for free and open access, it spread to major universities and research institutions across China within a few years, and even expanded overseas. But with the surge in visits, the cost of servers and maintenance also rose accordingly.
For a long time, Tang Jie spent every day modifying and debugging codes, and even hired an engineer specifically to maintain the operation of the system.
However, as a steady stream of free users poured in, looking at the rising expenses, Tang Jie once even thought about shutting down AMiner.
The turning point came from an overseas exchange. Foreign colleagues told Tang Jie that AMiner had saved them a lot of time looking for papers and peers. These words inspired him a lot, and he felt that this work was worth continuing.
Although users cannot do without AMiner, the operation still cannot do without capital.
Fortunately, soon some people came to talk about cooperation. The first to come were researchers with scientific research projects, who needed to screen out appropriate review experts from a large number of paper authors; then some local governments wanted to use AMiner to judge technology trends; there were also enterprises that wanted to find partners through papers and patents.
Therefore, Tang Jie developed AMiner into a free version and a paid version. The former continues to serve ordinary users, while the latter targets governments, scientific research institutions and technology companies, providing more in-depth and professional information and intelligence services.
Over several years, this free tool serving scholars gradually grew into a business targeting institutional clients, and a more mature team was formed.
Zhang Peng joined the AMiner team during this period. He was admitted to the Department of Computer Science and Technology of Tsinghua University in 1998, studied all the way from bachelor's, master's to doctorate, and stayed at the university to work after graduation.
Several other co-founders joined the team around the same time. Li Juanzi and Xu Bin have long been deeply engaged in knowledge graph, semantic computing and scientific and technological intelligence; Liu Debing graduated with a doctorate from the Institute of Computing Technology of the Chinese Academy of Sciences, worked at Technicolor China Research Institute before joining Tsinghua University; Wang Shaolan joined the team with years of experience in operating technology enterprises.
In June 2019, AMiner was officially spun off from the Tsinghua University laboratory, and Zhipu AI was established. Tang Jie, Liu Debing, Zhang Peng, Li Juanzi, Xu Bin and Wang Shaolan became the core founding team.
After the company was established, the division of labor was quickly clarified. Zhang Peng is responsible for system R&D and engineering implementation, turning the algorithms in the laboratory into products that customers can use for a long time; Li Juanzi and Xu Bin supplemented the knowledge graph research into the technical base; Liu Debing and Wang Shaolan took charge of financing, resource allocation and business decision-making.
At that time, although Zhipu AI had already started to make profits, it was essentially still an academic search engine, not the same species as the large models that were just emerging at that time.
However, the emergence of GPT-3 still made the team seriously think about one question: Can we compress the accumulation of the past more than ten years into a general large model, and integrate all the capabilities scattered in search, question answering, translation and code?
The problem was that compared with the track that was already making money, large models were full of uncertainties at that time. Whether to change direction and bet on something with extremely high cost and unknown result became the first major problem after the company was established.
Why Did They Choose the Most Difficult Path
After several discussions, Zhipu AI finally decided to change its direction.
What prompted them to make up their minds was the huge impact brought by ChatGPT. The team realized that the old business model of redeveloping algorithms every time for a new industry could be completely terminated by a certain general large model at any time.
In the past model, face recognition embedded in the security system required a set of algorithms, and knowledge graph embedded in government software required another set. Once the scenario changed, the products and algorithms all had to be redeveloped.
Large models, however, showed another possibility. As long as the model is large enough and the training data is sufficient, a single system can process different tasks such as question answering, writing, translation and code at the same time. This means that the previous practice of training a separate model for each scenario may be replaced by a general model for the first time.
For Zhipu AI, which had just been established for one year, this was both an attractive and dangerous direction.
Since AMiner could already generate revenue, continuing to do knowledge services meant the technology and customers were within their familiar scope. Training a general large model, however, required rebuilding algorithms, purchasing computing power, and recruiting talents, with investment far exceeding the affordability of a start-up company.
What's more troublesome is that all mainstream frameworks such as GPT, BERT and T5 are in the hands of overseas teams. Using the existing ready-made frameworks is low-cost and fast-acting; if they choose to build everything from the bottom up, they may spend several years and still fail to catch up with their competitors in the end.
Eventually, Zhipu AI chose the latter path.
After discussions, the team integrated the advantages of different pre-training methods and developed its own general language model framework — GLM. To put it simply, GPT mainly learns to continue writing according to the previous text, while GLM also needs to understand the masked context, and find a unified solution among continuation, fill-in-the-blank and semantic understanding.
This decision also redivided the roles of the founding team.
Among the six co-founders, Tang Jie controls the technical direction, Zhang Peng is responsible for turning the research results into an engineering system that can be trained, deployed and delivered, and Liu Debing serves as the chairman, coordinating strategy, resources and overall management.
It can be said that Zhipu AI's decisions from the very beginning were not issued by a single founder, but jointly promoted by this group of researchers and engineers who have run in with each other for a long time.
In 2021, Zhipu AI released the GLM-10B model with tens of billions of parameters. The next year, the team decided to increase the number of parameters to 130 billion.
The scale expanded by more than ten times, but the training difficulty was far more than ten times higher.
During the two full months of training, the hardware not only crashed randomly, but the model also collapsed without warning, and the video memory was always exhausted in unexpected places. Problems were still difficult to solve after repeated debugging.
Very often, due to the sudden interruption of training, the team could not even restore the last saved checkpoint, so they could only troubleshoot the data and code and start all over again.
These failures will never appear on the big screen at the press conference, but they test whether a company truly has the ability to train large models.
In 2022, the training of GLM-130B was officially completed and opened to global developers. In order to make this hundred-billion-parameter model truly usable, the team tried every means to make it run inference on GPU servers with limited configurations.
Opening weights and lowering the deployment threshold has also become part of Zhipu AI's technical route from this moment on.
In the same year, Zhipu AI also launched the code model CodeGeeX. Different from other application scenarios, the effect of code generation can be directly run and verified, which is not only efficient but also can be accurately measured. It was included in the company's product matrix very early.
But at that time, CodeGeeX was only a development auxiliary tool, far from being the commercial solution of the company. It was ChatGLM-6B, released in March 2023, that really brought GLM out of the scientific research circle and into ordinary people's computers.
This is a compressed open-source model with not the largest number of parameters, which can run locally on a consumer-grade graphics card. For Chinese developers who were just getting in touch with large models at that time, they did not need to wait for overseas interfaces or rent expensive servers, and a computer equipped with 3090 or 4090 could call the model.
With the usage threshold lowered, a large number of Chinese vertical models, personal experiments and enterprise prototypes emerged like mushrooms after rain. Free and open-source brought Zhipu AI a huge developer group, and also sent GLM into the test environments of more enterprises, along with new problems. The model can be downloaded for free, but the chips, power and engineers' time consumed in each training process are huge costs.
Therefore, after the transformation, Zhipu AI encountered a new problem: how to use an open model to support a large model company.
Losing 8.6 Billion Yuan in Four Years: How Zhipu AI Stays in the Game
For Zhipu AI, free is still a tried-and-tested strategy — open source the model first, get developers to use it, and then convert institutional customers who need local deployment into paid projects.
This actually hides the important reason why Zhipu AI insisted on developing the model from scratch in the first place.
For banks, operators, government departments and large manufacturing enterprises, when using large models in daily work, they usually do not casually open a public web page due to security considerations.
Moreover, since internal documents cannot be transmitted at will and data cannot leave the computer room, all models must be deployed locally and then connected to the existing business systems.
This kind of demand from public sectors is completely different from that of consumer products.
This means that Zhipu AI's early experience in serving scientific research institutions and government clients of large enterprises has become its exclusive advantage in the era of large models.
As a result, a steady stream of enterprises and public institutions came to seek cooperation on their own initiative. Compared with other large model companies, Zhipu AI is exceptionally clear about how to serve these institutions with unique demands when handling such requirements.
The most direct example is the cooperation with Postal Savings Bank of China in 2024. Zhipu AI connected the fine-tuned ChatGLM to the bank's anti-money laundering system, read transaction flows, customer identities and account information, extracted abnormal features and generated the first draft of reports, which were then reviewed by screening personnel. Throughout the whole process, the model was not only embedded into the bank's original risk control process, but also always stayed in the customer's own environment.
Local deployment has thus become Zhipu AI's most important source of revenue. In 2025, the company achieved a revenue of 724 million yuan, of which about 534 million yuan came from local deployment, accounting for 73.7%.
This means that for every 10 yuan Zhipu AI receives, more than 7 yuan comes from projects that need to be deployed on the customer's site.
But for Zhipu AI, although contracts from the public sector have large single amounts and high profits, each project is independent and difficult to replicate in batches. Because each customer needs to be received and served separately, revenue growth basically depends on how many projects the team can take on at the same time.
In contrast, cloud-based model services for the general public seem easier to scale up — customers pay per use, and the more usage, the more revenue. But the problem is that this path relies on "small profits but quick turnover", and domestic manufacturers have been fighting price wars, so the profit of each single call has been compressed to extremely low levels.
What's more troublesome is that while revenue keeps growing, the costs that need to be paid are also rising sharply.
From 2022 to 2025, Zhipu AI's total net loss in four years was about 8.608 billion yuan, close to 5.9 million yuan per day on average. In 2025 alone, R&D investment reached 3.18 billion yuan, while the revenue in that year was only 724 million yuan. This means that for every 1 yuan of revenue the company earns, the corresponding R&D expenditure exceeds 4 yuan.
The largest part of the entire R&D expense is still computing power. In 2024, Zhipu AI spent 1.553 billion yuan on purchasing computing power, nearly five times its revenue of 312 million yuan in that year.
There is no other way. Once the server starts training, it has to run continuously for several weeks. If there is a failure halfway, the consumed chip time and engineer working hours cannot be recovered.
This also explains why Zhip