First week of Z.ai and MiniMax’s IPOs
Financing for the labs: learning to sell products and scale.
Hi all,
First week back to work and I’m down with a nasty cold, so apologies for the late update. For this piece, I combed through 36kr, LatePost, bank reports, and spoke to investors and research analysts on the ground in Hong Kong. Hopefully, I can present a nuanced lens on these high-profile IPOs.
*I’m SO SORRY, for some reason, Substack did not save my latest line-edited version and blasted out a piece filled with typos. Anyway, you guys aren’t that unfamiliar with my jittery brain and typos by now. I cleaned things up somewhat.
First week of Z.ai and MiniMax’s IPOs
Two of China’s “foundation model six tigers” walked onto Hong Kong’s main stage within 48 hours last week. Zhipu, now branded Z.ai, went first, and the stock traded above HK$130, approximately 12% above the offer price.
MiniMax then demonstrated what a true retail frenzy looks like in this cycle: the public offer portion was only HK$220 million, yet it attracted over HK$253.3 billion in margin subscriptions, with 420,000 participants. By midday close, MiniMax was reported up over 78% to HK$294 from the offer price of HK$165, with a market cap of around HK$89.8 billion (~USD11.5 billion).
For many investors I spoke to, the thinking is actually quite simplistic. A rough heuristic is that if OpenAI is valued at ~USD800 billion and Anthropic at ~USD350 billion, a ~5–10 billion valuation for a “second-tier” company seems reasonable. Even if the probability is slim that one of these two becomes China’s best research lab, the chances are good enough to place a bet.
There’s no sugarcoating it: the business model is not fully clear, and the business is not that mature. Going public now is mostly for one reason: both companies have burned through too much cash and need to raise money quickly.
The thing is, Z.ai is still often classified under “lab learning to sell,” even though through their interviews, we’ve learned that the team was one of the first to think about what they call “P2P - paper to product and commercialization.
On the other hand, MiniMax has built a much more investor-friendly narrative as the company that already sells, because it has products that ordinary users can touch, or at least products that feel legible to non-technical investors.
A reputable independent Chinese business publication, 晚点LatePost, published two analytical pieces on the IPOs and argued that these listings are not a victory lap after the big showdown is won. Overall, LLM lab commercialization remains uncertain, but sustained R&D expenditure is assured, and the practical rationale for listing is to secure resources more efficiently. LatePost also noted that after Zhipu and MiniMax entered the secondary market back-to-back, which could help open the next chapter with larger private placements.
The listing is not the end of the story; rather, it is just the beginning. Now the two companies will have to build, pivot, make changes and reiterations, do R&D, and figure out a sustainable path forward, all under the public eye.
The best case scenario for them is to have raised money which would allow them to continue to push frontier research, prove to investors they have real potential to lead as a research lab in the mid-to-long term. The worst case… well let’s look at what happened to Figma, it went public, and within a year’s time it shot up to ~USD130 and then dropped back to roughly its listing price and never really made a comeback.
And that again shows how the public market is brutal. If they cannot continue to push boundaries and prove their worth, it will reward a narrative quickly, and it will also punish drift quickly. For both of these companies, the real test is not day one. It is whether the next four quarters look like cohesive strategic execution or like scrambling.
Anyway, for what its worth, this is the story of Z.ai and MiniMax and how they went from unknown labs to IPO darlings.
1. Z.ai: The lab before the company
When I interviewed Zixuan Li from Z.ai on Differentiated Understanding, he described Zhipu’s origin in a way that immediately explains why outsiders still treat it like a lab: “We originated from Tsinghua University’s data mining group, THUDM.” Under the THUDM name, he said the team became “very famous” in the open-source community for having “a lot of repos” and models, including CogVideo and CogView.
That open-source reflex has always been Zhipu’s strength and its branding problem. Even Zixuan himself wrote in a memo upon the IPO, saying, “But our global brand never really took off.”
The team of academics however were able to commercialize very early on. In an interview with CEO Zhang Peng and AI reporter Zhang Xiaojun, he said that the lab also groomed the founders of Megvii, one of the original AI four tigers. In 2016, the question was raised, how do we commercialize better and how do we push frontier and by then they have started thinking about the business of the science.
Over the years they’ve moved from ‘perceptual intelligence’ to ‘cognitive intelligence.’ Previous intelligence did not know what they do not know. But the work they are working on now is trying to really give the ‘intelligence’ cognitive ability, said Zhang Peng.
The company then built credibility model by model, through the developer community, but not the consumers. Some say it is because it is only a research lab, it’s too academic or technical. But Zixuan wrote on the day of its IPO, “Z.ai is a model company, but it has never been just a model company.” Credibility within the developer community is not the same as a successful product story. They only realized their need for this to be better understood by the wider public and other stakeholders in 2025.
The rationale behind that reflex is outlined in an internal letter that Zhipu authorized LatePost to publish around the IPO. Chairman Tang Jie opens with a small scene at HKUST: he runs into Prof. Yang Qiang in a coffee shop and jokes that he has been drinking too much coffee and needs to quit. Yang responds that being “addicted” is not necessarily bad if it makes you obsessed with research. Tang uses that moment to set the tone, then articulates the company’s long-standing North Star, “让机器像人一样思考” (“let machines think like humans”), and links the founding to a dual-system cognitive framing: “fast thinking + slow thinking.”
He is also blunt about the environment, which many seems to not want to admit. In the same letter, Tang is quoted as saying that since ChatGPT, “the industry has no real consensus, everyone is just moving forward.” If there is no consensus, you are underwriting a team’s ability to keep shipping despite confusion.
In many ways, you can see that the academics are actually true to their words without any glossing over details or trying to PR their way forward. Zixuan said during our conversation that for companies of their size, sometimes it’s just about trying to keep afloat and surviving. It’s hard to think about grander strategy when survival is at stake. And that candor is shared by Professor Tang Jie, who may be the one setting that tone in the company. He said at the most recent big AI conference in China that brought together Z.ai, Tencent, and many other leading AI labs - AGI-Next (which I’m grateful for Geopolitechs translating) that “The gap between Chinese and U.S. large models may not have narrowed,” a candid assessment of the technological gap.
2. 2021–2022: the company-building move that mattered most
The “lab to company” transition is usually described as a rebrand. Tang describes it as an operating decision.
In his letter, he writes that in 2020, they designed their GLM architecture and tried training a 10B base model, and that it was successful enough to be tried by enterprises, including Meituan. But he argues they were still far from AGI, and that the two missing pieces were the scale of knowledge and the ability to reason like humans.
Then comes the part that reads, in hindsight, like a pre-ChatGPT “all in.” Tang writes that in 2021–2022, large model development did not go smoothly, most people did not accept the “moonshot” plan, and the team still decided to train a 130B model. To keep the overall company rhythm intact, they set up two small innovation teams: one for training (later the “GLM three musketeers”), and one for building the MaaS platform. Tang says the two teams might not even have known the other existed.
Then in 2022, the company was able to train GLM-130B, the MaaS platform went online (bigmodel.cn), and Zhipu got its first cohort of real API users. Zhipu did not just build a bigger model. It built a surface where usage could be measured and, eventually, monetized.
However, in a 36Kr article on Z.ai and MiniMax’s IPOs, the period after late 2021 is a year when Zhipu, MiniMax, and their early investors explored and trialed “in silence,” and then “a year later, ChatGPT arrived, and everyone knows what happened next.” Before ChatGPT, investors kept asking why this mattered and how to commercialize “science.” After ChatGPT, the “why” became obvious, and the question moved to speed, distribution, pricing, and whether you can keep training.
But that doesn’t mean a foundation model company can’t be admired for its capability and still be questioned on monetization. The early investors, like Qiming Venture, were betting Zhipu would eventually find a repeatable commercial surface.
3. A business instinct earlier than outsiders assume
In the LatePost article, there’s an important detail that explains why Zhipu could move quickly once consensus arrived. The reporting says that as early as 2021, Zhipu had already cooperated with Alibaba Cloud and Sugon to provide model API services on the cloud. It also recalls a 2022 strategy meeting after a talk by Zhang Bo (then dean of Tsinghua’s AI institute), where they debated ToB versus ToC and concluded that ToC would be fiercely competitive; enterprises in China were already interested in applying large models, so Zhipu should lean B-side while keeping an eye on C-side. It adds that Zhipu’s compute reserve had reached “a thousand cards,” accumulated bit by bit by the founding team “begging” for resources.
In the internal memo, Tang describes the “DeepSeek shock” as hitting Zhipu’s ecological niche, in part because both companies had similar academically rooted research-team attributes and Zhipu had also contributed heavily to open-source ecosystems. The response was a “comprehensive return” to foundation model research.
Tang then turns that into a business cadence. In the internal letter, he writes that in 2025, Zhipu executed an explicit schedule: an April model to “steady the ship,” a mid-year model to “get to the table” as “one of the best,” and a year-end model aiming for “Top 1.”
He also anchors the model milestone in public scoring: the letter says GLM-4.7 launched and was open-sourced on Dec 23, and Artificial Analysis’s index ranked it first among domestic models and tied for global sixth with Claude 4.5 Sonnet.
4. Coding to monetize
On Differentiated Understanding, Zixuan’s most interesting commercial observation was blunt: “API users are not sticky.” He compared developer model choice to consumer subscription behavior and argued Z.ai does not want users to pay purely by tokens. “We want them to pay by value or by the product itself,” he said, explaining why they chose a subscription-like GLM Coding Plan.
Tang’s letter gives the numbers that make this feel like more than a pricing experiment. He writes that the GLM Coding Plan is used by 150,000 developers across 184 countries, and that after GLM-4.7, the MaaS platform’s annualized ARR exceeded RMB 500 million, including over RMB 200 million overseas, rising from RMB 20 million to RMB 500 million in about 10 months.
Then he makes the strategic logic explicit: 2025 was hard, there was a price war, model results early in the year were not meeting expectations, and they needed a precise breakthrough. “We finally found Coding as the breakthrough,” Tang writes, describing GLM-4.1 as an early probe and GLM-4.5 as a decisive push.
And during the AGI-Next conference, he revealed that “At the beginning of 2025, we had intense internal debates. In the end, we decided to bet on coding. We went all in—and it turns out, we made the right call.”
5. Global expansion, but to the “south”
Zixuan told me Zhipu has been pushing beyond China, especially toward the Global South. Tang’s letter gives a concrete example of what “global” looks like in their language: Malaysia’s national-level MaaS platform is built on Z.ai open-source models, and GLM is described as becoming a “national-level” model there.
The prospectus showed that overseas revenue is still early but measurable. Overseas revenue was 11.6% of total revenue in 1H25 and is likely to continue to grow.
Tang’s 2026 roadmap is explicit and technical: GLM-5, new model architecture designs beyond Transformer constraints, more general RL paradigms, and continual learning, so that deployed models are not static.
He also announced X-Lab, a new internal department meant to gather younger talent for frontier exploration and incubation across new architectures and cognitive paradigms, including projects beyond software if needed. Now, we’ll need to wait and see how it all plays out in the next 6-12 months.
MiniMax: founders, funding, strategy, frontier bets, then an IPO
1. Founding story
MiniMax was founded in early 2022, right before the ChatGPT wave, and the reporting is explicit that Hillhouse was the first investor. The founding pair is Yan Junjie and Yun Yeyi, and the reporting is explicit that they were colleagues at SenseTime before starting the company.
Yan was born in 1989 and grew up in a county town in Henan, earned his PhD at the Chinese Academy of Sciences Institute of Automation, and spent seven years at SenseTime, rising from researcher to its youngest VP, including running smart city and gaming businesses. Yun is a Johns Hopkins graduate and previously led strategy in SenseTime’s CEO office.
2. Seven rounds of fundraising
In seven pre-IPO rounds, 30 institutions invested US$1.5 billion into MiniMax. Alibaba invested the most money; Hillhouse led the first round and, by stake, was second only to Alibaba; and Future Capital (明势资本) participated in the most rounds, according to the prospectus.
In the LatePost article, it’s said that Hillhouse partner Li Liang spoke with Yan and Yun for three hours in its early days, then produced a term sheet with the valuation line left blank and asked them to fill in the valuation and investment amount. Yan’s plan was US$30 million at a US$200 million valuation.
The same reporting describes the “conviction” behind that decision: a Hillhouse investor said Yan delivered a nine-hour “technical class,” covering Transformer scaling laws, GPT-3 progress, DeepMind reinforcement learning, diffusion models, and CLIP, and that few people could connect those dots coherently at the time.
Then after that, it was recognized as a likely leader in the Chinese LLM space. Yunqi and IDG joined quickly, and the angel round raised about US$31 million at a US$200 million valuation, close to the founders’ plan, and MiniMax declined offers to take more money at a higher valuation.
Future Capital becomes the recurring character. The reporting says that before ChatGPT, MiniMax had two rounds, and the second round’s only new shareholder was Future Capital. It also says Future invested in 6 of 8 rounds if you include the IPO cornerstone placement and the seven pre-IPO rounds.
3. The burn and the breakthroughs
The company started off with strong 2C products, unlike Z.ai. One of the earliest “proof” was Glow, an app that allowed users to create AI characters and storylines. The reporting says MiniMax released Glow in October 2022, did very little paid acquisition, and accumulated over one million ACG users in two months. Then ChatGPT launched in November, and Glow became “a small ripple.”
Then the funding wave came. The reporting says MiniMax raised US$260 million in early 2023 at a post-money valuation of US$1.157 billion, with Tencent, Xiaomi, and Xiaohongshu among strategic investors, and Shunwei and Oasis among new shareholders.
There’s also a discipline signal embedded in the same package. It says Tencent wanted to invest more in that round, and MiniMax ultimately took only US$50 million from Tencent.
It’s reported that MiniMax put almost all R&D resources into MoE models in 2H23 and failed twice in training. In 2024, it put 80% of resources into a linear attention architecture model, later released as M1 in early 2025.
Ultimately, the business breakdown seems more mature and the revenue breakdown behind the product portfolio is that Talkie/Xingye, Hailuo AI and MiniMax Voice, and the Open Platform API business each contributed about 30% of revenue, a roughly balanced 1:1:1 mix.
4. Now an IPO
First I must preface, I have less material on MiniMax since I haven’t spoken to them directly (disclosure: my former firm is the lead consulting firm for their IPO- but I was not on that account) so my analysis of them comes from publicly available information and view of views on them. But I’m not the only one who thinks the IPO was more out of necessity to survive for them - actually for both Z.ai and MiniMax. LatePost wrote that commercialization is uncertain, sustained R&D investment is certain, and IPO is a more efficient way to obtain resources. China internet analysts agreed with this sentiment when I met them.
On the morning of the listing, the reporting quotes Yan saying he hopes MiniMax can contribute to raising the industry’s intelligence level and that they had “initially explored” a grassroots path for AI entrepreneurship, even though the road ahead remains challenging.
What the IPO changes for both companies
The reality is that these companies are now public inside a category where the next training run is not optional. The two tigers have celebrated their IPOs; in many ways, we’ve said that, it was out of pure necessity as they were both burning through cash and out of money to compete in this capital-intensive race.
LatePost and 36Kr both essentially have said the same thing in different ways: the speed of AI iteration and the speed of resource burn force higher financing demand, and by the end of 2024, most of the “six tigers” already had IPOs on their agenda because the primary market couldn’t keep supporting the valuation ladder forever. The reason why I wanted to showcase what the analysis from within China is to show that it’s not like people domestically are just celebrating this without being mindful of constraints.
What the actual consensus and review show that is even shareholders “regret” that cash on both balance sheets still isn’t enough for globalization, and both companies are still searching for a deeper pool of capital. Public markets don’t just fund capability. They fund cadence. Cadence is what turns research into an operating plan.
There is also a shared post-IPO management problem that rarely gets discussed in model narratives: talent. The reporting says listing helps retain people, and notes that after DeepSeek’s release, Zhipu granted equity incentives to some employees, which internal staff viewed as stabilizing morale. It quotes an investor saying IPO turns incentives from “drawing pies” into real benefits, and that ByteDance and Alibaba will intensify competition, so companies need to give employees confidence or reasons and incentives to not leave for bigger labs that are more well-funded.
The backdrop goes back to the fact is that venture capital is dry in China at least in relative terms to the U.S. That was even one consideration for Manus to go offshore to start with, probably. Furthermore, as Chinese regulators turn to hinder Manus’s sale to Meta, this may lead to more Chinese founders being cautious about actually building domestically. If capital and exits are restrained, then well, pragmatically, why stay?
And for Z.ai and MiniMax - now the hard part begins: doing all of the things we just wrote about, while being public.
public references
HKEx news (prospectuses and filings): https://www.hkexnews.hk
MiniMax website (prospectus also published here per HKEx announcement): https://www.minimaxi.com
HK eIPO application portal (mentioned in MiniMax HKEx announcement): https://www.hkeipo.hk
Forbes (Yan Junjie): https://www.forbes.com/sites/ywang/2026/01/09/founder-of-chinese-ai-model-developer-minimax-becomes-a-billionaire-as-shares-surge-on-listing/
Z.ai Zixuan Li thread (Z.ai 2025: Fueling the Path to AGI)
晚点LatePost:MiniMax 融资故事:4 年 7 轮,谁在推动中国 AI 第一场资本盛宴
晚点LatePost:晚点独家丨智谱上市,唐杰内部信要求全面回归基础模型研究
AI Proem / Differentiated Understanding: Ep. 19 with Zixuan Li
张小珺商业访谈录:129 全球大模型第一股的上市访谈,和智谱CEO张鹏聊:敢问路在何方?






Thank you. Good updates.