Bonus Exclusive | Zhang Fan, COO of Zhipu, is about to leave. Will Zhipu be the next SenseTime?
36Kr "Career Bonus" (ID: ZhiChangHongLi)
The commercialization issues that large model companies are frequently questioned about are becoming increasingly prominent. Zhipu AI, which has targeted OpenAI since its inception and aimed its commercialization direction at the B - end and G - end, is now facing such a difficult problem.
Exclusive information obtained by "Career Bonus" reveals that Zhang Fan, the COO of Zhipu AI [1] will leave the company at the end of June to continue his entrepreneurship in the field of AI Agent. At the beginning of this year, there was a change in Zhipu's commercialization department: a small - scale "personnel flow" took place. Although Zhang Peng and Zhang Fan still jointly responsible for commercialization, the businesses under their management were reorganized, and the specific division method no longer followed the ToB/ToG logic.
Regarding Zhang Fan's departure for entrepreneurship, "Career Bonus" verified the information with Zhipu AI. Its public relations department responded: "Zhang Fan's new entrepreneurial project after leaving is also a part of Zhipu's MaaS (Model as a Service) platform ecosystem, helping Zhipu's models complete the last - mile transformation. His new project has received investment support from Zhipu. Currently, Zhipu regards training the next - generation base model as its top - priority strategy, focusing on continuously improving model performance, outputting model intelligence through the MaaS platform, expanding commercialization, and focusing on cultivating the commercial application ecosystem and innovative Agent ecosystem of the MaaS platform."
In addition to the personnel changes on the commercialization side, a source pointed out that Dong Yuxiao [2] the head of Zhipu AI Institute, left the company at the end of last year, and Zhipu AI Institute has changed from overall technical planning by large teams to small - scale special project research led by doctoral students in each small unit. However, this statement was denied by Zhipu's public relations department, which said that the former's account still exists in the internal communication software.
"Career Bonus" learned that several core personnel within Zhipu AI were previously Tang Jie, Dong Yuxiao, Zhang Peng, Zhang Fan, and Zhang Kuo. Combining the previous public personnel information, we have made a more detailed summary of key figures in the following figure.
Summary of key figures
According to "Intelligent Emergence", Zhang Kuo, the VP in charge of financing at Zhipu, left the company at the end of January this year. A person familiar with the matter told "Career Bonus" that after Zhang Kuo left, it is currently difficult for Zhipu to obtain new financing in the market. According to public information, since 2025, Zhipu has received a total of 1.8 billion yuan in strategic investment from state - owned assets in Hangzhou, Zhuhai, and Chengdu.
Since last year, the departure of Zhipu AI's core personnel has slowed down both the model capabilities and financing plans of Zhipu. How will Zhipu and the other AI "Six Little Dragons" break the deadlock in commercialization in the future?
[1] Zhang Fan is a serial entrepreneur in the field of intelligent technology, with more than ten years of industrial implementation experience in intelligent technology in multiple fields such as automotive, tourism, and enterprise services. He once served as the CTO of Souche Group, founded Yuanyin Intelligence and served as its CEO, founded Miaoji Travel and served as its CEO, and also engaged in technical management work at ByteDance, Tencent, and Sogou.
[2] Dong Yuxiao is an associate professor at the Department of Computer Science, Tsinghua University, and a member of the Knowledge Engineering Laboratory (KEG). His main research directions include data mining, graph machine learning, pre - trained models, and social networks.
Questions about commercialization, the difficulty of ToB MaaS ╱ 01
Departure of core personnel, slow progress in model capabilities ╱ 02
The next SenseTime? ╱ 03
Questions about commercialization, the difficulty of ToB MaaS
Exclusive information obtained by "Career Bonus" shows that at the beginning of this year, there was a change in Zhipu's commercialization department: a small - scale "personnel flow" took place. Although Zhang Peng and Zhang Fan still jointly responsible for commercialization, the businesses under their management were reorganized, and the specific division method no longer followed the ToB/ToG logic:
CEO Zhang Peng manages some businesses + branch companies;
COO Zhang Fan is in charge of some businesses + regional branch companies.
Relevant people analyzed that such an adjustment might be to facilitate the implementation of local government projects and ensure delivery efficiency.
The B - end market in the AI field was once regarded by insiders as the key to forming a closed - loop income. Zhipu aimed at the domestic B - end market from its inception.
Before the business reorganization, Zhang Fan was in charge of Zhipu's B - end business, and Zhang Peng was responsible for government, military, and overseas businesses. An insider told "Career Bonus" that after Zhang Fan left, Zhipu basically had to abandon the large - scale enterprise services and focus on government projects. And the latter might directly lead to human resources taking on project customization. "The more projects are won, the more losses may be incurred."
However, in the AI field, it has become a consensus that the domestic B - end [3] market is difficult to operate. Last year, DeepSeek initiated an API price war, so that the revenue of large - model service providers from "small B" orders was not much, and they all turned to the "big B" market.
On the one hand, the real demand from domestic clients is decreasing. An insider said: "Now for many projects from enterprises, customers often don't pay because of demand but because of operations and investment promotion." If the products and solutions of large - model service providers do not address the pain points of enterprises, not only will the budget given by enterprises be limited, but also the project requirements are often unclear, which will lead to frequent rework of projects, extended project duration, and increased labor costs.
In terms of R & D cost investment, an insider told "Career Bonus" that Zhipu has now grown to a scale of 800 - 1000 people, and the number of employees in the commercialization team accounts for half of the total. However, the official denied this statement, saying that "R & D accounts for more than 70%, and there are also many professional R & D personnel in the business departments."
On the other hand, there is also involution among large - model service providers.
A typical phenomenon is: In order to get orders, model service providers can only "make exaggerated promises" on the one hand and lower prices on the other hand, but the quality of the projects is often unsatisfactory. "Career Bonus" learned from an insider that in 2024, another unicorn enterprise won a bidding project at a low price. However, at the project implementation stage, it was found that the enterprise's technology was not sufficient to meet the project's requirements, and then Zhipu took over the project.
When the gross profit margin is not clear to the outside world, an insider said, "Currently, there is a risk of long payment cycles and even bad debts in the bills." However, a Zhipu employee said that he did not notice this phenomenon. According to a report from "Caixin Magazine", Zhipu AI's revenue reached 300 million yuan in 2024, but the loss in the same period was as high as about 2 billion yuan.
With the release of DeepSeek - R1 during the Spring Festival this year, Zhipu's commercialization space has been further squeezed. "Career Bonus" learned from an employee of another large - B service provider that now the customers of customized projects are also inclined to use DeepSeek's models, which makes it even more difficult for enterprises like Zhipu to get orders.
[3] There are two types of B - end orders in the market: big B and small B. Big B generally refers to that the service provider needs to undertake large - scale customized orders. The advantage of this kind of order is that the project turnover is high, ranging from hundreds of thousands to millions of yuan. However, the disadvantages of big B orders are also obvious. The service provider company needs to provide customized services, with high costs. Small B orders refer to large - scale enterprise services, such as API and cloud services provided by service providers. The advantage of small B orders is that they are easy to scale up and have low costs, while the disadvantage is that the order turnover is often not high.
Departure of core personnel, slow progress in model capabilities
One of the reasons for the blocked commercialization is also the stage - by - stage bottleneck of technology.
The most intuitive manifestation is that the last update of Zhipu AI's basic large model stopped in December 2024 when the deep - reasoning model GLM - Zero - Preview was released. From the beginning of this year to now, Zhipu has not released any new models.
In the ever - changing AI field, a cruel fact is: Not moving forward is almost equivalent to falling behind.
According to SuperCLUE evaluation data, whether it is Zhipu's basic model performance, reasoning model performance, or open - source model capabilities, it is not among the top, and even lags behind DarkSide and StepStar, which are also part of the "Six AI Little Dragons".
● The result form of SuperCLUE's evaluation of the capabilities of various models
● The result form of SuperCLUE's evaluation of the capabilities of various models
Before the release of DeepSeek - R1, the discussion about "whether large models should be open - source or closed - source" was intense. In April 2024, Li Yanhong, the founder of Baidu, publicly stated "Open - source models will become more and more backward." Zhipu, DarkSide, and StepStar were all supporters of closed - source models.
The release of DeepSeek - R1 shattered the "illusion" of closed - source models: DeepSeek - R1 not only trained a model comparable to GPT - 4 in terms of capabilities at a very low cost, but also open - sourced the technical details and was open - sourced under the MIT license, allowing developers to use, modify, and commercialize it freely without additional authorization.
The open - sourcing of a stronger model and the ability of developers to directly use it have made the advantages of closed - source models disappear. The once "Six AI Little Dragons" were all in a state of confusion and entered a "period of confusion".
From the external performance, large - model manufacturers this year have, on the one hand, focused their promotion on products; on the other hand, the progress of manufacturers in the technical field has become more "low - key", and they are "forced" to open - source some of their models.
In 2025, Zhipu's only new move was to release the open - source GLM - 4 - 32B - 0414 series of models. StepStar open - sourced the voice interaction model Step - Audio and video generation model Step - Video - T2V developed in cooperation with Geely Automobile Group. Baidu plans to open - source its ERNIE Bot 4.5 series on June 30, 2025.
An insider told "Career Bonus" that in 2025, large - model unicorn enterprises are facing a new round of financing pressure, and they dare not be too aggressive in technological investment; on the other hand, it will take several months, as well as human, material, and financial resources to catch up with DeepSeek. With the upcoming release of R2, everyone is waiting.
"Zhipu still attaches great importance to technology internally," an internal employee told "Career Bonus". "Everyone is anxious. Recently, Zhipu has been recruiting a large number of algorithm talents to catch up with Alibaba's Tongyi Qianwen and DeepSeek."
The next SenseTime?
In April 2025, Zhipu AI officially completed the IPO counseling record filing with the Beijing Securities Regulatory Bureau, becoming the first enterprise among the "Six Little Dragons of Large Models" to start the listing process.
Among the former "Four AI Little Dragons", SenseTime was the first to achieve an IPO. As a "pioneer" in the traditional AI field, which originated from academia and focused on customized projects, there is an interesting comparison between SenseTime and Zhipu.
Will Zhipu be the next SenseTime?
Both enterprises were led by influential professors in academia and the industry in the early stage: SenseTime was led by Professor Tang Xiao'ou from the Chinese University of Hong Kong, and Zhipu AI was helmed by Professor Tang Jie from the Department of Computer Science, Tsinghua University. At the beginning of their establishment, the core members of both SenseTime and Zhipu were mainly composed of the students and teachers of the two professors.
However, there is no natural connection between academic glory and business success.
The unique temperament of academic - oriented enterprises once showed contradictions in SenseTime. "Career Bonus" learned from an insider that due to its emphasis on academia, SenseTime was once called a "paper factory" - the entire enterprise was obsessed with research and publishing papers, resulting in that the large - model product "SenseChat" was not released after 10 months of internal testing, simply because "no one cared about it if it was released and no papers could be published."
This "emphasizing research over implementation" tendency exposes the deep - seated conflict between academic and business logics.
Some Zhipu employees also told "Career Bonus" that Zhipu treats employees from Tsinghua University and those recruited from the society completely differently: "Employees from Tsinghua University eat self - heating hot pots from Haidilao and drink Perrier, while ordinary employees only have cheap snacks purchased by the enterprise at a low price per catty on the shopping platform."
It seems that in order to avoid being criticized for the characteristics of the "academic school", Zhipu chose another path: actively embracing commercialization from the very beginning. Zhang Peng, the CEO of Zhipu AI, once said that the company entered the market with technology and customers from the start, understood the domestic To B market, and knew what kind of products to use for B - end commercialization.
However, the same is that the commercialization of both companies is aimed at high - cost customized services. The difference is that in the early stage, SenseTime's gross profit margin reached 56.5%.
