Manus was acquired, and Zhipu is set to go public in 8 days.

```

Big AI news, one after another.

This morning, news just came that Manus was acquired by Meta, and soon after, the race around the “world’s first AI large model IPO” also heard the sound of a shoe dropping.

On December 30, Beijing Zhipu HuaZhang Technology Co., Ltd. (hereinafter referred to as “Zhipu”) officially launched its Hong Kong IPO. The subscription will last until January 5, 2026, and it is planned to be listed on the main board of the Hong Kong Stock Exchange on January 8, 2026, with stock code “2513”.

According to the IPO arrangements, Zhipu intends to offer a total of 37,419,500 H shares globally, including 1,871,000 shares in the Hong Kong public offering and 35,548,500 shares in the international offering.

The IPO pricing and fundraising scale were also revealed—each share will be priced at HK$116.20. After deducting related issuance expenses, the expected fundraising amount is about HK$4.3 billion, corresponding to an estimated IPO market value exceeding HK$51.1 billion.

Public information shows that Zhipu’s cumulative financing in the private market has reached RMB 8.344 billion, and its latest valuation has climbed to RMB 24.377 billion. This means that in this key leap towards listing, Zhipu’s market valuation has almost doubled, making such a premium listing a significant market challenge.

The cornerstone investor lineup is also quite eye-catching. Announcements show that cornerstone investors intend to subscribe for a total of HK$2.98 billion, accounting for nearly 70% of this issuance (assuming the over-allotment option is not exercised).

Institutions participating in the cornerstone subscription include: JSC International Investment Fund SPC, JinYi Capital Multi-Strategy Fund SPC, Perseveranc(e) Asset Management, Shanghai GaoYi Asset Management, WT Asset Management, Taikang Life, GF Fund, 3W Fund Management, and 11 other investment institutions.

Against the backdrop of overall pressure on Hong Kong tech assets, such a high proportion of cornerstone subscriptions also writes a clearer market note for this race for the “world’s first large model stock”.

The cash burn continues, large models are entering the capital market

Looking inside the industry, the six AI large model startups that were highly touted in 2024 have already shown clear divergence: two chose to withdraw from foundational model competition, instead focusing on vertical applications.

The other four—Zhipu, MiniMax, Moonshot AI, and StepStar—are still trying to stay at the large model table. In late December 2024, Zhipu and MiniMax successively filed their prospectuses with the Hong Kong Stock Exchange.

Unlike MiniMax, which focuses on the consumer market (to C), Zhipu mainly focuses on enterprise solutions (to B), already having applications in finance, internet, smart devices, healthcare and other industries.

In the first half of this year, Zhipu’s revenue was RMB 191 million, while the loss for the period reached RMB 2.358 billion. AI R&D expenses were as high as RMB 1.595 billion.

If in 2023, the high valuations given by the primary market to large model startups were more of a bet on grand technological narratives, then entering 2024–2025, the market has clearly shifted towards evaluating model capabilities and commercialization pathways.

Even leading companies are hard-pressed to avoid continuous investment in foundational models, and startups must also directly face challenges such as whether models can be continuously iterated and applied to real scenarios.

And a lot of this depends on whether the capital market is willing to provide long-term, stable funding support.

In April this year, Zhipu began A-share IPO counseling with the Beijing Securities Regulatory Bureau. But as of December 12, the company had not received further opinions or inquiries from the China Securities Regulatory Commission regarding its A-share listing.

Against this background, Zhipu chose to turn to the Hong Kong stock market to seek more sustainable fuel for this high-investment, long-cycle large model race. At the same time, it will directly face the dual tests of financing capability and market confidence—whether anyone is willing to pay for AI’s long-term investment.

From GLM to MaaS: Zhipu’s large model technology foundation and commercialization path

According to the prospectus, Zhipu mainly provides services ranging from computing power, API interfaces to MaaS (Model-as-a-Service), supporting both local and cloud deployment, and has been implemented in multiple industries.

As one of the representative companies engaged in general language model research and industrialization in China, Zhipu’s technology system is centered on GLM, covering text, multimodal, and application-oriented model services.

GLM is a large language model paradigm based on Transformer, unifying modeling for both understanding and generation tasks by combining autoregressive generation and masked prediction. The architecture was first proposed by Zhipu and the relevant Tsinghua University research team and has been iteratively improved in subsequent models.

In 2021, Zhipu released China’s first proprietary pre-trained large model framework GLM, and launched a MaaS (Model-as-a-Service) product development and commercialization platform, providing large model capabilities and services to external clients via the platform.

In 2022, Zhipu released and open-sourced GLM-130B (a bilingual Chinese-English 100+ billion parameter model), marking the formal use of the GLM system for pre-trained large language models.

In January 2024, the GLM series achieved an important milestone with the launch of GLM-4, which supports longer contexts and faster inference, greatly reducing inference costs.

In July 2025, Zhipu further open-sourced GLM-4.5. The model topped the Hugging Face (the world’s largest open-source model platform) trending chart globally within 48 hours of its launch.

In September of the same year, Zhipu released and open-sourced GLM-4.6, a further upgrade of the foundational model, mainly enhancing encoding capabilities. In November, GLM-4.6 ranked first globally on CodeArena.

In December, Zhipu launched the latest flagship model GLM-4.7:

• In core coding, compared to the previous generation GLM-4.6, GLM-4.7 made significant improvements in multilingual agent programming and terminal-based tasks: SWE-bench 73.8% (+5.8%), SWE-bench Multilingual 66.7% (+12.9%).
• Ambient programming: GLM-4.7 made an important leap in UI generation quality, able to create simpler, more modern web interfaces, and provide more accurate layout and size control in presentation generation, improving overall visual effects.
• Tool usage: GLM-4.7’s tool use capabilities improved significantly and demonstrated stronger practical operational abilities in web browsing tasks covered by BrowseComp.
• Complex reasoning: Achieved 42.8% on the HLE (Humanity’s Last Exam) benchmark, up 12.4 percentage points from GLM-4.6.

Compared with GPT-5, GPT-5.1-High, Claude Sonnet 4.5, Gemini 3.0 Pro, DeepSeek-V3.2, Kimi K2 Thinking, GLM-4.7 has also performed well:

Meanwhile, Zhipu has also launched multimodal models for different functions, including CogView (image generation), GLM-4.5V (visual understanding and reasoning), CogVideoX (video generation), and others.

For AI Agents, Zhipu’s foundational agent model is AutoGLM. In December, Zhipu fully open-sourced the core AutoGLM model, marking AutoGLM’s further development in the open ecosystem.

As of June 30, 2025, Zhipu models have supported more than 8,000 institutional clients; as of the last practical date, they have supported about 80 million devices.

On the commercialization side, Zhipu began laying out its MaaS business model back in 2021.

The MaaS platform offers four main categories of model capabilities, covering language models, multimodal models, agent models, and code models, and also provides an integrated toolchain supporting model fine-tuning, deployment, and agent development.

From model capability expansion, to advancing agent technology, to the gradual formation of the MaaS commercial system, Zhipu has completed a relatively comprehensive round of technology and product deployment.

But the shoe dropping does not mean the endgame is set. As the company goes public, high R&D spending, rising compute costs, and the as-yet-unvalidated general large model commercialization path will all be exposed to more transparent scrutiny.

IPO is not the end, but rather a longer cycle of public testing.

Source:Synced (Jiqizhixin)

Risk Warning and DisclaimerThe market has risks, investment needs to be cautious. This article does not constitute personal investment advice, nor does it take into account the special investment objectives, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their particular situation. Investment based on this is at your own risk. ```