Less than an hour after GLM 5.3 was released, a Z.ai salesperson’s phone and WeChat were lighting up. More than a dozen customers wanted to know when the API would go live. Others, recalling the scramble around Kimi K3, wanted to lock in Z.ai’s inference compute in advance.
It went on all weekend.
According to 36Kr, GLM 5.3 set a new high for Chinese models in coding at launch. It tied Moonshot AI’s Kimi K3 for first place among open-source models and ranked in the same performance tier as proprietary flagship models including Claude Fable 5 and GPT-5.6 Sol.
Customers responded immediately.
The weekend ended, but the messages and calls did not. The GLM 5.3 API, originally scheduled to go live on August 18, had been delayed, and the sales team was inundated again.
“A customer was yelling at me on the phone: ‘We’ve got the internal budget approved, and development and operations are all ready. Every day of delay costs us tens of thousands more in management expenses!’” the salesperson said.
The rush had been building since around February, the salesperson said, when Z.ai released GLM 5. The subsequent launch of GLM 5.1 moved the company closer to the global top tier, while GLM 5.2 later emerged as the state-of-the-art, or SOTA, open-source model on several benchmarks.
Each version represented a material improvement.
The contrast with the earlier Z.ai, then known as Zhipu AI, was substantial.
When large language models took off in China in 2023, Z.ai was an awkward fit for the prevailing narrative. Founded in 2019, it did not resemble the young, consumer-oriented artificial intelligence startups attracting much of the attention.
Nor did it look like a conventional commercial AI company. It emerged largely from Tsinghua University and retained a distinctly academic, laboratory-like culture.
Its commercialization path was also unfashionable by industry standards. For years, Z.ai focused on private model deployments for government and enterprise customers rather than the scalable, replicable products favored by internet companies.
Yet by 2026, the company looked very different.
Z.ai became the world’s first publicly listed foundation model maker. At one point, its market capitalization reached HKD 1.3 trillion (USD 165.8 billion). Its open-source models gained international recognition, while people familiar with its finances told 36Kr that it had become the largest Chinese model developer by annual recurring revenue (ARR).
Coding has been central to that transformation.
As foundation model companies around the world moved into the coding market popularized by Anthropic, Z.ai was arguably among the first Chinese companies to establish a meaningful position. Its coding models helped lift its reputation and revenue and intensified competition among China’s large language model developers.
Understanding why Z.ai moved into coding early, and how it gained ground, requires going back to May 2025, two months before the release of GLM 4.5.
A reset and an unexpected breakthrough
In May 2025, Z.ai convened an emergency strategy meeting.
“Only the core executives and a small number of shareholders attended. One of the central questions was where Z.ai’s next generation of models should go,” a person familiar with the matter told 36Kr.
The stakes were high.
DeepSeek R1 had broken out during the 2025 Lunar New Year holiday. Its performance was comparable with OpenAI o1, while its price was about one-thirtieth as high.
That threatened the economics of Z.ai’s commercialization business.
DeepSeek R1 disrupted the enterprise model market at a time when customized deployments for business customers remained Z.ai’s most important source of commercial revenue.
One Z.ai employee responsible for commercial delivery told 36Kr that the Lunar New Year break had been anything but restful. “I was taking five or six customer calls a day, and most of them were asking whether we could deploy DeepSeek,” he said.
He had mixed feelings about it. “DeepSeek became a household name overnight. A lot of customers’ bosses were telling their teams to switch to DeepSeek.”
Another Z.ai employee told 36Kr that, by a conservative estimate, nearly 30% of the company’s customers shifted to DeepSeek at the time. To retain some of them, the employee said, Z.ai offered steep discounts.
China’s foundation model market was also changing.
Interest in chatbots was fading. Moonshot AI and MiniMax, often grouped with Z.ai and several other leading Chinese AI companies, were betting on flagship models designed to use tools and execute complex tasks.
Faced with that pressure, Z.ai’s strategy meeting produced a decision that would reshape its model development.
The company abandoned its previous approach of building separate vertical models for text, multimodal, coding, and other capabilities. Instead, it decided to build a single model with a large parameter count, trained on a combination of reasoning, coding, and agentic data.
Z.ai had concluded that it needed to move beyond its existing strengths. Its next opportunity, management believed, would come from models capable of solving more complex productivity problems.
The decision was contested.
“Quite a few shareholders opposed continuing to scale up the model because the investment required was too high,” one person familiar with the matter told 36Kr.
Z.ai had long been under financial pressure. Its 2025 financial report showed a net loss of RMB 4.718 billion (USD 700.4 million). R&D expenses alone reached RMB 3.182 billion (USD 472.4 million), about four times its revenue for the year.
There was another problem. A model trained by combining multiple types of data did not fit neatly with Z.ai’s enterprise customization business.
A Z.ai employee working on its B2B operations said the company had previously divided its model capabilities into clear vertical categories to match specific customer use cases. Conversation, image generation, video generation, and coding each had corresponding models and products.
Using one integrated model across those scenarios would make deployment more expensive.
That three-in-one model became GLM 4.5, which launched in July 2025.
“It was originally supposed to launch in April, but because the model direction changed at the last minute, it was pushed all the way to July,” a person familiar with the matter told 36Kr.
For Z.ai, much depended on the result.
The company prepared 15 trillion tokens of general-purpose data and another eight trillion tokens of coding, reasoning, and agentic data to train the model. Its total training data volume was nearly 1.5 times that of comparable models at the time, according to 36Kr.
Pressure spread across the company. Several Z.ai employees told 36Kr that, in the months before GLM 4.5 launched, algorithm researchers at its R&D institute practically lived at the office.
“The day [GLM 4.5] was released, I left work in the middle of the night,” one employee said. “When I came back in the morning, the algorithm team was sitting there in exactly the same formation. It looked almost identical to when I’d left the night before.”
Marketing employees also worked through the night on launch day, in some cases until 8 a.m.
GLM 4.5 ultimately became Z.ai’s first model to build a strong reputation for coding. It also put the company among the Chinese model developers that moved earliest into the emerging market.
Z.ai had found a potential new growth engine: coding.
Several Chinese foundation model companies had noticed signs of the opportunity as early as June 2024, when Claude 3.5 Sonnet was released. But many were reluctant to commit.
MiniMax founder Yan Junjie once asked DeepSeek founder Liang Wenfeng two years ago: “Are you going to do AI coding?”
Liang said no.
“The consensus at the time was that there might be only one to two million people in China who could write code, and that didn’t seem like a large enough market,” said Yan at an event.
What the industry had underestimated was how AI coding could expand the market itself. If AI tools materially changed how software work was done, a market serving two million programmers could eventually reach many more users.
Still, Z.ai’s early success was not entirely planned.
A Z.ai employee said that when training of the three-in-one model began, reasoning, coding, and agentic capabilities did not have sharply differentiated priorities.
One reason coding later became the focus was user demand.
According to 36Kr, Z.ai’s R&D team repeatedly discussed the need to stay close to real-world user needs rather than optimize primarily for benchmark rankings. Actual model capabilities, the team believed, had to be validated through practical tasks.
The algorithm team therefore relied heavily on feedback from users and customers, and one request surfaced repeatedly: improving R&D productivity.
That also aligned with management’s emphasis on pushing the upper limit of model intelligence.
After GLM 4.5 launched, one salesperson recalled a steady stream of requests from key accounts seeking to integrate the model into programmers’ workflows.
On social media, some users began describing GLM 4.5 as a lower-cost alternative to Claude. Its coding performance was approaching Claude Sonnet 4 on some evaluations, while its price was about one-seventh as high.
Coding became GLM 4.5’s most popular capability.
In September 2025, Z.ai launched the GLM Coding Plan, which the company positioned as the first coding plan offered by a Chinese foundation model developer.
An enterprise edition followed in October. That same month, Z.ai launched GLM 4.6, which it said had been specifically strengthened for coding.
By then, the company’s new commercial direction was becoming clearer.
Alongside its labor-intensive and difficult-to-scale enterprise deployment business, Z.ai had found a coding business with greater potential for scale.
“The chatbot story was already over once DeepSeek arrived. What we should be thinking about is what the next bet is,” Z.ai founder Tang Jie said at an event in early 2025.
That bet turned out to be coding.
GLM 5.2 and a surge in market value
When Z.ai listed in Hong Kong in January this year, becoming the world’s first publicly traded foundation model company, CEO Zhang Peng sent every employee an RMB 200 (USD 29.7) red packet.
The celebration lasted one day.
A day later, MiniMax went public. Its shares surged 109% on the first day of trading, pushing its market capitalization above HKD 100 billion (USD 12.8 billion), nearly twice Z.ai’s.
For the first time, two foundation model companies with very different cultures were being judged side by side in the public market.
Z.ai was initially valued less favorably.
“At the time, MiniMax’s focus on the consumer market was seen in China as having more room for imagination,” an investor in foundation model companies told 36Kr. “Z.ai, with its B2B and government business labels, just didn’t look as sexy.”
Five months later, the market narrative shifted again after Z.ai released another open-source model that reached the top of several benchmark rankings.
The change was broader than it had been with GLM 4.5. Revenue, market attention, and Z.ai’s share price all rose sharply.
Then on June 13, Z.ai released its next-generation flagship model, GLM 5.2.
Coding remained its hallmark. On Artificial Analysis’ overall leaderboard, GLM 5.2 ranked as the leading open-source model, behind only proprietary models Claude Fable 5, Claude Opus 4.8, and GPT-5.5.
Customer spending followed.
GLM 5.2 drove rapid growth in Z.ai’s API revenue. A Z.ai shareholder told 36Kr that in May, the company’s ARR was around USD 500–600 million, with an expectation of reaching USD 1.0–1.5 billion by year’s end.
One month after GLM 5.2 launched, 36Kr reported that Z.ai’s ARR had reached USD 1 billion by July, nearly doubling in two months. Its end-of-year ARR forecast had also been raised to USD 2.5 billion.
Z.ai did not respond to a request for comment on the figures.
Cloud providers that earn a share of revenue from selling third-party models quickly responded to demand for GLM 5.2.
One Z.ai employee told 36Kr that Alibaba Cloud, Volcano Engine, Xiaomi, and Kingsoft Cloud were promoting GLM 5.2 heavily to customers.
An Alibaba Cloud salesperson confirmed the trend to 36Kr. Because downstream demand was strong, by July Alibaba Cloud was “directly recommending GLM 5.2 to customers.”
For Z.ai, the release became an inflection point reminiscent of DeepSeek’s earlier breakout. Its impact also extended beyond China.
According to monitoring by cloud deployment and hosting platform Vercel, GLM 5.2’s usage grew faster than that of any other model released in 2026, surpassing DeepSeek V4, which launched in April.
“Z.ai doesn’t have much of a marketing budget,” a salesperson told 36Kr. “What we do is give potential key accounts and opinion leaders early access for free, so they can experience the model’s capabilities directly.”
A marketing professional also told 36Kr that before GLM 5.2 launched, Z.ai’s middle and senior managers traveled overseas whenever they had the opportunity to meet key accounts and technology key opinion leaders, inviting them to participate in early testing.
“A lot of overseas customers tried GLM 5.2 and found that its overall capabilities were better than DeepSeek V4,” the person said. “That’s how the overseas reputation turned around.”
Z.ai’s share price rose alongside the model’s reception. On the first trading day after GLM 5.2’s release, its shares jumped about 32%. One week later, its market capitalization crossed HKD 1 trillion (USD 127.6 billion).
Even before GLM 5.2, as its share price climbed earlier in the year, Z.ai had made an important commercial decision: sharply increase the priority of its model-as-a-service business.
Some shareholders told 36Kr they believed Z.ai’s original customized enterprise business alone could not support a much larger valuation.
Customization offered limited economies of scale. To take on more contracts, Z.ai had to hire more people.
According to one employee’s calculation, the enterprise business had been growing at about 1.5 times annually. That was steady growth, but the model had an obvious ceiling.
Every contract was also highly bespoke.
“Sometimes, just to get the acceptance signoff as quickly as possible, we had to put in a lot of extra emotional and physical labor outside our normal work,” an employee said. “We’d attend meetings on the customer’s behalf, help them prepare presentation decks, and even pick up their kids. Put simply, it took a huge amount of manpower to keep customers happy.”
The pace of model development created another problem.
AI models now improve so quickly that one model may lead the industry when a deployment project begins, only to lag competitors by the time implementation is completed several months later.
That has made MaaS offerings more attractive to customers seeking flexibility.
Z.ai had discussed such a business model as early as 2021.
Even then, people inside the company believed it should take cues from OpenAI and place greater emphasis on charging for API usage.
But China’s MaaS market was still nascent, and domestic models were not yet competitive enough to support the strategy.
Five years later, Zhang Peng has chosen Anthropic as Z.ai’s benchmark.
The gap remains substantial.
According to 36Kr, Z.ai’s price-to-sales ratio is more than five times Anthropic’s. If Z.ai traded at the same price-to-sales ratio, it would need an ARR of HKD 48.78 billion (USD 6.2 billion), or about USD 6.22 billion, to support a market capitalization of HKD 1 trillion. Its ARR had only recently reached USD 1 billion.
Tradeoffs and avoiding major mistakes
After GLM 4.5 established Z.ai in coding and GLM 5.2 accelerated its commercial growth, employees began hearing the same question more often. Why Z.ai?
When probed, several employees gave strikingly similar answers: disciplined focus, avoiding major mistakes, and making the right technical choices at key stages.
“If you get every step right, the model can hardly turn out bad,” one person said.
Some employees pointed to ByteDance’s success with Seedance as another example of the same principle.
That relatively conservative approach is closely tied to Z.ai’s culture.
Looking back at several of its critical decisions, one pattern stands out: the company repeatedly concentrated on areas where its technical strengths were most relevant.
Zhang Peng once described Z.ai this way:
“It’s like a male engineering student at Tsinghua. Very smart, very capable. Whatever you ask him to do, he can do it very well. But he just doesn’t provide much emotional value.”
The description reflects how outsiders have often viewed the company: technically strong and strategically perceptive, but less naturally suited to consumer products.
There is also a parallel between Zhang’s comment about Z.ai lacking “emotional value” and the behavior of some of its models.
In 2024, Z.ai worked with a hardware manufacturer to integrate GLM into a product. During internal testing, when an employee prompted, “I’m feeling a little down,” the GLM-powered product responded: “Turn on the air conditioner.”
Zhang has previously said in a media interview: “Zhilin, the founder of Moonshot AI, knows how to understand ordinary people’s needs and thoughts. We may not be as good at that. It has to do with our positioning.”
That weakness may have put Z.ai at a disadvantage in consumer-facing products.
Two examples are its chatbot ChatGLM and video generation model Qingying.
According to AICPB, ChatGLM had 10.43 million monthly active app users as of March 2025. Doubao had 97.36 million during the same period.
Qingying also struggled commercially, according to a Z.ai salesperson.
“Most of the customers are people working in artistic fields, and they don’t think much of Qingying’s aesthetics,” a Z.ai salesperson told 36Kr.
Rather than continue allocating resources evenly across those areas, Z.ai increasingly concentrated them where it believed it had an advantage.
“Z.ai is a company with a very engineering-oriented character,” one industry participant told 36Kr. “Coding is the direction that fits it best.”
At Z.ai’s R&D institute today, the algorithm team responsible for multimodal model training has been reduced to around ten people. The teams focused on text and coding, by contrast, number around 100.
A person familiar with the matter said Z.ai has allocated substantial budgets to building reinforcement learning environments and data engineering for coding and agentic capabilities.
Given constraints on compute, model capacity, and training data, developers must choose which capabilities to prioritize.
Trying to optimize every capability at once can leave a model without a clear strength.
“Every post-training objective takes capacity away from other objectives,” a former Z.ai algorithm researcher said. “Making coding ability a training objective means the training data needs to be concise and highly precise, with less unnecessary chatter. That runs against the kind of emotionally rich communication people normally have with one another.”
The employee cited GPT-5 and DeepSeek V4 as examples. When both models were released, some users complained that their “emotional intelligence” had declined as their coding and reasoning capabilities improved.
MiniMax offers another example. After the M3 model was released, an algorithm researcher who participated in its training said none of its capabilities stood out because its training objectives had failed to converge.
“MiniMax’s strongest businesses are all consumer-facing. The bosses knew coding and agentic capabilities were important, but they still couldn’t let go of the work on optimizing the model’s emotional intelligence.”
The result was a muted industry response to M3.
More recently, however, 36Kr learned that MiniMax disbanded the “general entertainment group” within its post-training team at the end of June and redirected people and resources toward reasoning and workplace use cases.
Z.ai’s ability to make tradeoffs was only part of the story. Employees also pointed to its system for reducing the risk of major errors during model training.
A finely tuned A/B testing process is one example.
“When there’s a key technical choice to make, Z.ai will split people into two or three groups to run A/B tests,” one employee told 36Kr. “Each test group has two or three researchers. It’s run like a laboratory.”
Another employee said Z.ai breaks the training process and technical routes down in considerable detail. When the algorithm team cannot decide which route to take, it evaluates each against a test set and selects the best-performing option.
Experience is another factor.
As early as 2021, Tang Jie led the training of WuDao 2.0, a trillion-parameter foundation model initiated by the Beijing Academy of Artificial Intelligence.
Among the founders of China’s major AI model companies, Tang and Moonshot AI founder Yang Zhilin were among the few with experience working on models at that scale. People familiar with Z.ai said that experience helped shape the company’s approach to data and post-training.
Strong coding performance today depends heavily on supervised fine-tuning, or SFT, and agentic reinforcement learning, or RL, conducted in environments designed around agent tasks.
In both stages, the quality and composition of data can matter as much as the algorithms themselves. Designing an effective data strategy, however, requires experience.
“Which queries do you use? Which tasks do you run? Should the synthetic data be used for SFT or RL? Should it be used to generate simulated training environments, or used directly for training? None of those decisions can be separated from experience,” a MiniMax algorithm researcher told 36Kr.
After M3 launched, MiniMax held an internal postmortem. One conclusion, according to the researcher, was that its data strategy had not been good enough.
Z.ai, meanwhile, had already invested in an overseas data procurement team in 2024.
Internally, Zhang Xiaohan, who is responsible for model training and data, led the development of a data labeling platform called Kuangbiao.
During internal rotations, teams waiting for their next training cycle have one primary task: labeling data.
Since GLM 4.5, the algorithm team has also placed greater weight on feedback from downstream customers.
One Z.ai employee recalled that the company’s R&D institute frequently met with leaders from the sales and delivery teams to collect feedback from key accounts.
Before a model was released, it had to pass two internal evaluations. The first took place inside the institute and tested general capabilities. The second involved the delivery team, which tested the model against high-frequency scenarios used by target customers.
“That’s why GLM has such a strong reputation for coding among developers,” the employee said.
The talent moat
Behind Z.ai’s system for reducing errors are hundreds of algorithm researchers at its R&D institute.
“You can’t poach Z.ai’s algorithm talent with money,” a headhunter specializing in senior hires told 36Kr.
“Big technology companies have offered compensation packages worth tens of millions of RMB to recruit core algorithm researchers from Z.ai’s research institute, and still failed.”
One reason is compensation.
As Z.ai’s market capitalization rose, stock options held by some key researchers came to be worth more than RMB 1 billion (USD 148.5 million), according to people familiar with the company.
But employees and recruiters said another factor was the network of academic and personal relationships connecting Z.ai with its core talent. Those ties are difficult for competitors to replicate.
Much of the company, from its founder to employees across key roles, has roots at Tsinghua University. Many are connected through shared academic advisers, classmates, or teacher-student relationships.
“It’s a combination of a company, a university, and a research institute,” one investor said.
For an organization centered on model training, one person is especially important: founder and chief scientist Tang Jie.
A professor in Tsinghua University’s Department of Computer Science and Technology, Tang was among the earlier scholars in China to work on machine learning and natural language processing.
Several sources told 36Kr that when they were looking in late 2022 for Chinese companies that could become equivalents of OpenAI, the first Tsinghua academics who came to mind were Tang Jie, Sun Maosong, and Huang Minlie.
Many Z.ai employees describe Tang as enthusiastic, persuasive, and execution-oriented. They also said he sets demanding standards for both technology and business.
One algorithm researcher at Z.ai’s R&D institute said that when training results for its agent product AutoGLM 2.0 were unsatisfactory, Tang postponed its launch.
The company’s core team is also connected through Tsinghua.
CEO Zhang Peng, who runs commercialization, earned his bachelor’s, master’s, and doctoral degrees at the university and, like Tang, comes from its computer science department.
Z.ai president Wang Shaolan, who is responsible for government relations, and chairman Liu Debing also graduated from Tsinghua University. Both previously worked with members of the founding team at the university’s Big Data Research Center.
More than 70% of the hundreds of researchers at Z.ai are current or former undergraduate, master’s, or doctoral students from Tsinghua’s computer science department, according to 36Kr.
That concentration of teachers, graduates, and students has created a relatively stable algorithm organization shaped by adviser-student relationships and a strong academic culture.
It is visible in the way the R&D institute works. The institute has a flat hierarchy and operates more like a laboratory than a conventional corporate organization, according to employees. Important directions often begin with a technical question or a discussion about an experimental result. Researchers from different backgrounds join the discussion, while younger researchers have considerable room to challenge assumptions and test ideas.
Many students told 36Kr they would be reluctant to leave because of the respect and personal attachment they feel toward their teachers and senior schoolmates.
“Some students who want to pursue master’s or doctoral degrees under Tang will choose to work at Z.ai first,” a former Z.ai algorithm researcher said.
A steady supply of new talent, combined with relatively low turnover, has become one of Z.ai’s organizational advantages.
According to 36Kr, after GLM 5 launched, one Z.ai algorithm intern received nearly ten calls from headhunters in a single day.
ByteDance recruiters have gone through author lists on GLM research papers and contacted researchers individually.
Tencent’s Hunyuan team was willing to offer one Z.ai master’s graduate twice the base salary to move, according to 36Kr.
On August 19, Tang Jie posted on X to explain his view of the scaling law. Model capabilities, he argued, are no longer determined simply by parameter count. Instead, model development increasingly resembles a controlled experiment in which the variables include how much data is used, where the compute budget is spent, and who ultimately runs the model and under what conditions.
That argument helps explain the work behind GLM 5.3.
Compared with its predecessor, the model’s gains came primarily from post-training rather than simply increasing its pretraining parameter count, according to 36Kr.
Other leading foundation model companies, meanwhile, have continued increasing model size. Over the past month, Moonshot AI, DeepSeek, and Alibaba Group have all released new flagship models, with parameter counts ranging from 160–280 million, two to four times that of GLM 5.3.
Among them, Kimi K3, released by Moonshot AI on July 17, displaced GLM 5.2 after the latter had held the open-source SOTA position for about a month.
Just over 20 days later, Z.ai reclaimed the top position with GLM 5.3.
The broader race remains unsettled. Model direction, training routes, data quality, and algorithm talent all continue to shape how foundation model companies compete.
For Z.ai, coding was the capability that aligned those pieces at the right time.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu and Zhang Yuxin for 36Kr.
Note: HKD, RMB figures are converted to USD at rates of HKD 7.84 = USD 1 and RMB 6.74 = USD 1 based on estimates as of September 1, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.