The Rise of GLM and China’s Zhipu AI

0:00 In early 2026, two AI labs, Minimax and Zai, IPO'd on the Hong Kong Stock Exchange. While Minimax is interesting today, I want to look at the second of the two. Zai is more well known in China as Zhipu and they produce the well known GLM series of open weight models. If DeepSeq is China's OpenAI, then would that make Zhipu more like it's anthropic? In today's video, we profiled one of China's top AI labs. This video is brought to you by the asianometry Patreon after doing his undergraduate at the prestigious Tsinghua University in 2002, Zhang Peng joined Tsinghua's Knowledge Engineering Group, or KEG, as a master's student.

0:46 KEG focuses on knowledge graphs and information extraction. They've been around since 1996, when they started out doing Semantic Web development, data processing and domestic Java projects there. Zhang, an introvert from Hubei who likes spicy and rarely shows his anger, worked with esteemed professors like Tang Jie, a distinguished computer science professor originally from Sichuan. In 2006, Tang Jie's KG team helped launch a project called ArnetMiner, or just a Miner, a scientific and technology research intelligence mining program running on a single desktop computer.

1:22 The concept sounds rather lofty, but it's more like Google Scholar. It builds a knowledge graph and ranking of researchers on the web that we can search. Machine learning was used to extract data like emails or researcher profiles out of unstructured web profiles to build the network. As the project matured, Zhangpang started to think about commercialization. The KEG culture heavily focused on bringing actual projects to market. They had a phrase called P2P, or paper to product.

1:53 Faculty members encouraged it, saying that AI is inseparable from industry. In 2015, the a miner team started to earn money with partnerships with private companies. There was a prominent one with Microsoft that involved data sharing and academic exchange, but others included Elsevier, Google, Exxon Mobile and IBM. The US Dollars the research team earned from these partnerships were super encouraging, and they started thinking about fully commercializing a miner by starting a company around it.

2:26 Several lab members had already left to start companies of their own. Notably, Tang Wenbing and yang Mu founded Megv. Megv is a facial recognition AI company that is one of mainland China's first generation of AI dragons, worth $4 to $5 billion today. They are known for their face API. In an interview with the Chinese podcaster Zhang Xiaozheng, which is fantastic and informs much of this early history Zhang Peng recalled initial excitement in 2015 over computer vision.

2:58 However, there was also spirited debate over whether it can deliver general artificial intelligence. But with people unclear on the definition of AGI still the case today, Zhang decided to focus on an intermediate level. He labeled cognitive intelligence. Simply put, its goal is to get machines to think like humans. By 2016, Zhang Peng and his co founders Tang Jie, Liu De Bing, Li Zhengzi, Shi Bin and Wang Xiaolan decided to turn a miner into its own company.

3:33 However, it took time until commercial companies can be formed around certain university cooked products. In 2018, China's Department of Education issued guidance that allowed the commercialization of this certain work done at research institutions. After a long negotiation, the team finally formally founded Zhipu, or Knowledge Atlas. In June 2019, Tsinghua University transferred over the a minor product and remains one of zhipu's larger shareholders.

4:03 Zhang started off as the cto, running the engineering teams and delivering products to customers. Tang Li and Xu worked on the research side, writing papers and the like. Liu helped with negotiating with the university and Ip and Wang served as the company's founding CEO. But later Zhang took over sometime in late 2022 and Wang thusly moved up to the president role. Back in 2018, OpenAI released the first version of GPT GPT1.

4:33 Since it did not work that well, most did not pay attention. The second version, GPT2, started to gain some prominence within academic circles. Then in 2020, OpenAI released GPT3, which scaled up the parameter size to 175 billion and became shockingly good at generating human like almost text. Zhang recognized it as a breakthrough. Traditional natural language processing ran sentences through a complicated pipeline to identify a sentence's structure, tag its nouns and verbs, and break it down into a tree.

5:09 While it worked for some sentences, it was brittle and did not scale. GPT3 showed that none of that had been necessary, and the results were astonishing in how it outperformed current cutting edge models like Google's Bert Zhang Peng recalled reaching out to his senior at KEG, Professor Zhang Bu to discuss GPT3. Zhang Bo, no relation agreed that while the LLM represented a new paradigm, it also spouted total nonsense. In other words, it didn't know what it knew.

5:39 Nevertheless, Zang, with backing from debingliu, pushed his colleagues to train their own multi billion parameter bilingual model, supposedly saying unless you take the leap and make the pursuit, you'll never know just how fast or far you can run. End quote. Since OpenAI didn't release many details on how GPT3 worked they had to figure it out on their own. In late 2020, co founder Tang Jie worked with 100 scientists from several mainland Chinese universities on a government backed project called Wudao.

6:11 Wu Dao sought to train ultra large models. In March 2021, Wudao 1.0 announced their progress, presenting four big models. There was Wenyuan, a 2.6 billion parameter model with GPT3 like performance in Chinese NLP tasks like answering questions and grammar. There were also specialty models like Wenzu, a biology oriented LLM, and Wenlan, which connected languages to images using 50 million image captions scraped off the web.

6:43 The fourth model and most special was Wenhui, which the group said was the first novel pre trained model approaching cognition. With a little fine tuning, the 11.5 billion parameter model can do complex reasoning, generate images and even write poetry. The Wenhui model was trained with a pre training framework that Tang Jie called General language model or GLM. It was first introduced in March 2021 in a paper originally titled All NLP Tasks are Generation Tasks, but a year later it was resubmitted as just glm.

7:16 I suppose for some rebranding. Anyway, GLM grew out of work done to find a compelling alternative to prior models like GPT and bert. While they are all transformers, they work differently. Famously, GPT generates text token by token, going from left to right. It's unidirectional, predicting the next token based on what has come before. Google's BERT works differently. It blanks out the middle of a text and then generates the token based on the blank's surrounding context. That makes it bi directional the B in bert.

7:48 Some have called it a fill in the blank bot. GLM blends the two together. It fills in a blank using the surrounding context, but does it token by token? Dang said that it was a distinctive approach that was more stable during the training process. And the team noticed that beyond 100 billion parameters, the model began to show unexpected capabilities in understanding human input. Zhang Peng cited this one example of a simple a newborn baby's head is 1/4 of its body length.

8:16 If a baby is 50 centimeters long at birth, how long is its head? It's not a trick question, don't overthink it. Anyway, Zhang noted that the model did not calculate the answer by straightforwardly multiplying 50 by 1/4. Rather, it divided by 4, and Zhang saw that as a sign that the model actually understood the intent behind the question. Wudao was not explicitly a ZHIPU project, but it was led by co founder Tang Jie. And of course GLM would give its name to the company's series of foundation models.

8:50 As Depu worked on glm, it produced other AI products via a system integrator ISH set setup. This led to a bunch of rather wide ranging projects in addition to the original A Minor Jhipu's other projects included Xiaoke, a physical robot that interacted with people in libraries and museums, and Hua Zhibing, which got some attention as mainland China's first virtual AI student to enroll at Tsinghua University. Hua was powered by a larger model called Wudao 2.0 at 1.75 trillion parameters, said to be China's largest and started class in June 2021.

9:28 Under Tang Jie. Her image and voice were generated by an AI during the Winter Olympics. The company also contributed to a Digital Sign Language broadcasting system. Zhang attributes this broad spectrum of projects to its Monetization framework model as a service, which it launched in 2021. The idea was that we can deploy large models to customers via cloud, API or local deployment for those who want on premise. It sounds slick, but it also reminds me of the AI system integrator business model that mainland China's older AI Dragons Megv, SenseTime, etc.

10:07 Took on years before. I think those companies struggling to make real impact dampen investor interest in future startups bearing AI breakthroughs. In December 2021, Zhang and his research team at the Keg Lab in Tsinghua finally resolved to go for their next big model. While Wudao 2.0 had 1.7 trillion parameters, it was a sparse mixture of experts model, meaning that only a percentage of the model activates to process a query, so the model does not perform at its full 1.7 trillion parameter count.

10:42 By comparison, a dense model like GPT3 uses all of its parameters and they were seen as having a simpler architecture, more consistent quality output and easier to deploy. Zhang said they wanted to make something that can be run on a single A100 GPU server. The training run began in January 2022 and they struggled through a variety of problems like random hardware failures, model gradients blowing up, unexpected memory usage and more.

11:13 Jephu contributed more compute and in May 2022 the final successful training run kicked off on 96 A100 Nvidia servers and it was completed after two months in July 2022. The total cost of compute was 4 million RMB with a total all in cost of 10 million RMB 1.5 million USD. Then, after overcoming all these problems, they open sourced the model. Why? Zhang said that if they kept it closed, then they couldn't have earned any money off it immediately and they wanted to share it.

11:48 Open source was the spirit of the culture in a pre ChatGPT environment, and it did indeed seem to garner good reputation for Chinese AI efforts. In November 2022, on the eve of ChatGPT, Stanford University released a report benchmarking 30 prominent models from 12 providers, including the GPT models. Zhang was proud that Tsinghua's GLM130B was the only Chinese model to be benchmarked, and while the authors noted that it underperformed closed models, they did say that it represented the progress that open models had made.

12:24 That being said, Zhang recalled that investors then were not all that impressed. The Zhipu's team had trained a GPT3 like Model 1, citing the weak economy as straight up to cut the valuation in half. This all changed on November 30, 2022 when ChatGPT came out. When the ChatGPT moment came, Zhang was in Shenzhen raising money and securing customer contracts. He remembered not being too surprised that a Chat product like ChatGPT was coming.

12:54 OpenAI was still publishing at the time and they also employed a good number of Tsinghua alumni that the guys at Zhipu talked to. So there was a heads up, but I don't think anyone expected the thing to blow up like it did. For some 10 years, these AI researchers had worked in obscurity. Now all of a sudden, it was coming into vogue. Everyone had caught AI fever. Tang Jie held an online lecture titled Reflections on the journey from 100 billion parameter models to ChatGPT.

13:24 In late December 2022, over 1000 Tsinghua alumni attended, asking questions about data, applications and robotics. For Zhang, this was a validation of their direction. As one of the few Chinese companies with their own LLM models, ZHIPU suddenly became very popular. Zhang Fan, the company's former coo, recalled in a podcast, being totally overwhelmed by messages. Internally, zhipu worked through the Lunar New Year holiday to finish their own chatbot.

13:55 At the same time, they prepared a smaller LLM to be open sourced ChatGLM6B, the idea being to allow Chinese using ChatGPT to poke around inside a real life model. Zhang Pen recalled 2023 being full of excitement and anxiety. Excitement because he saw this huge wave of opportunity and anxiety because Mainland China in 2023 experienced a Titanic influx of new AI models. At the start of the year There were maybe four or five GPT3 class LLMs in addition to Baidu, Huawei, ByteDance and Alibaba's R&D branch called Damo Academy.

14:37 While Advanced, all of them were a generation behind GPT 3.5. The next few months saw a massive number of new startups and models be announced in the market, a chaotic mad dash known as the Battle of the Hundred models. Some of mainland China's top AI startups were founded in 2023. This includes Baituan AI, founded in April by the former CEO of Sogo, a search engine company backed by Tenxing. I think today they're focused on the healthcare space.

15:09 Moonshot AI, which is a startup behind the Kimme Chatbot Co founder and CEO Yang Zhilin studied under Tang Jie at Tsinghua. Founded in March, they started sending out beta invitations for Kimme in October 2023.01ai, which is the AI startup founded by former head of Microsoft Research China and Google China Kai Fu Lee. They raised a large round early on, but not sure what they're up to today. And of course there is Deepseek Hedge fund billionaire Liang Wenfeng formally founded the Chinese blue whale in July 2023.

15:43 DeepSeek shipped their first coding model in November and people quickly recognized its Legitness. A September 2023 survey done by a brokerage found over 130 models available in mainland China, not counting variants for specific industry verticals like health or finance. Even then, it was clear that a shakeout was imminent, that things would soon descend into a price war with how insane 2023 was. Zhang Peng says that he sometimes wonders how he and Zhipu managed to get through it all.

16:15 In 2023, they announced the first of their foundational chat models, ChatGLM. ChatGLM 2 and ChatGLM 3 quickly followed. They also opened Source a smaller version of the model called ChatGLM6B, which became quite popular in the open source community at the time. They also raised a boatload of money after their series A. In 2022, they did a series B1 through to B6 round of financing with B2 B3.1 and B3.2.

16:46 In between, Zhipu launched their consumer chat product Zhipu Qingyuan and I think the name means elegant words. In August 2023, despite being among the first chatbots to hit the mainland Chinese market, it fell behind the rest of its cohort like Baidu's ErniBot and ByteDance's Doubao. Those firms had powerful pre existing distribution channels. Doubao would eventually become mainland China's leading chatbot with an estimated 324 million monthly average users.

17:19 Moonshot's Kimmy Deepseek, Tencent's Yuan Bao, and Alibaba's Quinn later joined the battle too, spending large sums via red envelopes or whatever to acquire customers. Even at its peak, qingyuan never broke 11 million MAU. Zhipu tried buying users for it, but ordinary Chinese consumers were and still are not very willing to spend money on premium chatbot features. And as a bunch of academics they they probably felt themselves not temperamentally suited for a bloody rock fight inside the consumer octagon.

17:53 Which is probably why Zhibu ended up doubling down on the enterprise with their model as a service business, which I mentioned earlier. Zhang explained that while excessive price cutting does happen in enterprise, it happens less than in consumer because in the end the client wants the product and is suspicious of a bid that is priced by too low. Zhipu began its rise above the foray in January 2024 when they released GLM4, the 200 billion parameter GLM4.

18:23 They dropped the chat to indicate its general intelligence launched with agentic capability tools like Python, web browsing and image generation. And interestingly, it was a closed model accessible to developer customers only via API. Evaluations at the time estimated that it was about 90% of GPT4 with outperformance, particularly in Chinese language tasks and alignment, though a few commenters did grumble about its closed source nature as well as the censorship.

18:53 A few Months later, in May 2024, DeepSeek launched their V2 model with radically lower pricing. Jhipu, in response, cut prices for their GLM3 Turbo API. However, Alibaba Cloud, Baidu AI Cloud and Tencent followed suit, triggering the price war so many had anticipated late the previous year. 2025 opened with a bang. First, the United States added the company to the entity list, presumably due to links with the Chinese military. The entity list placement limits the company's access to certain US Origin software tools and other products.

19:28 A week later, the Deepsea crash hit the markets. The startup's new R1 model caused a lot of people to temporarily rethink the value of Nvidia stock and a lot of other people to Google what Jevons Paradox is Deepseak's sudden worldwide prominence influenced Jepu and the direction they eventually took on their next models before Zhang recalled that people doubted whether a mixture of expert structure could beat GPT4 level. But DeepSeek showed the way. Zifu studied the model and made their own modifications, leading to the release of their First MOE model, GLM 4.5 on July 28, 2025.

20:08 Also for the first time, they open sourced a Frontier model. 4.5 was open. Wait. At first people were mostly like oh wow, another open source Chinese model. But 4.5's quality did surprise several observers. It reaches first place in Hugging Face's rankings within 48 hours. Then other companies, American ones, even start offering it in their products. Zhang realized then that new models don't need big flashy announcements anymore. So Zhipu keeps releasing new models throughout the rest of 2025, dropping them with a simple post on the socials.

20:44 GLM 4.6 and 4.7 are well received for their coding ability, which people say are quite good considering the low cost. With revenue growing at roughly 100 to 150% a year, albeit off a low base, the company goes public in January 2026 and of course the stock goes bonkers. Soon after the listing, the stock price soared some 10 times. In June 2026 the company's peak worth was $170 billion, more than Xiaomi or BYD and more than half of Alibaba.

21:15 That is of course way too high and the market cap has since settled down to something like 50 billion as of this writing. In July they sold shares to institutional investors to raise more money and I think that is not going to be the last time they do that. The company has continued to release very well regarded models. GLM 5.3 and its flash version in particular seems to have gathered a lot of attention, though Deepseek remains the de facto number one Chinese AI lab.

21:44 In September, Jepu reported six month revenue to be up by 400% to $140 million. Their ARR is about $1.6 billion on a monthly basis in the midst of switching revenue models from being largely on premise model deployment to cloud API, which is perceived as being more scalable. But the company's biggest money issue is that it spends more than twice the amount in R and D as revenue, likely due to all the compute for training new models.

22:13 Then they give those models away for free. How long can this immensely dilutive business model go on? We shall see. Alright everyone, that's it for tonight. Thanks for watching. Subscribe to the Channel, sign up for the Patreon and I'll see you guys next time.