Major developments are unfolding rapidly in the AI sector.

Shortly after the morning announcement that Manus had been acquired by Meta, another significant event concluded the race for the title of "the world's first major large language model stock."

On December 30, Beijing Knowledge Atlas Technology Co., Ltd. (hereinafter referred to as "KNOWLEDGE ATLAS") officially launched its Hong Kong IPO offering. The subscription period will run until January 5, 2026, with a planned listing on the Main Board of the Hong Kong Stock Exchange under the stock code "2513" on January 8, 2026.

According to the offering details, KNOWLEDGE ATLAS plans a global offering of 37.4195 million H shares, comprising 1.871 million H shares for the Hong Kong public offering and 35.5485 million H shares for the international offering.

The IPO pricing and fundraising scale were also revealed—the issue price is set at HK$116.20 per share. After deducting relevant offering expenses, the expected net proceeds are approximately HK$4.3 billion, corresponding to an expected IPO market capitalization exceeding HK$51.1 billion.

Public information shows that KNOWLEDGE ATLAS has raised a cumulative total of 8.344 billion yuan in the private market, with its latest valuation climbing to 24.377 billion yuan. This indicates that in its crucial leap towards listing, the company's market capitalization has nearly doubled. Such a substantial "premium listing" also presents a significant market challenge.

The lineup of cornerstone investors is notably impressive. The announcement reveals that cornerstone investors have collectively agreed to subscribe for HK$2.98 billion, accounting for nearly 70% of the total offering size (assuming the over-allotment option is not exercised).

The institutions participating as cornerstone investors include 11 investment firms such as JSC International Investment Fund SPC, JinYi Capital Multi-Strategy Fund SPC, Perseveranc Asset Management, Shanghai Gaoyi Asset Management, WT Asset Management, Taikang Life Insurance, GF Fund Management, and 3W Fund Management.

Against the backdrop of overall pressure on Hong Kong-listed technology assets, such a high proportion of cornerstone subscription provides a clearer market endorsement for this race to become the "world's first major large language model stock."

The cash burn continues as large language models begin their journey into the capital markets.

Within the industry, a clear divergence has emerged among the "Six Tigers" of AI large model startups, which were highly sought after in 2024: two have chosen to voluntarily withdraw from the competition in foundational models, shifting their focus instead to vertical applications.

The remaining four—KNOWLEDGE ATLAS, MiniMax, Moonshot AI, and StepFun—are still attempting to remain at the table in the large model arena. In mid-to-late December 2024, KNOWLEDGE ATLAS and MiniMax successively disclosed their Hong Kong IPO prospectuses.

Unlike MiniMax, which focuses on consumer-facing (to C) applications, KNOWLEDGE ATLAS primarily concentrates on enterprise-level solutions (to B), which have already been deployed in sectors including financial services, internet, smart devices, and healthcare.

In the first half of this year, KNOWLEDGE ATLAS reported revenue of 191 million yuan, while its loss for the period amounted to a substantial 2.358 billion yuan, with AI R&D costs reaching 1.595 billion yuan.

If the high valuations granted to large model startups by the primary market in 2023 were largely a bet on grand technological narratives, then by 2024-2025, the market began shifting more explicitly towards evaluating model capabilities and paths to commercial realization.

Even leading companies cannot avoid the need for continuous investment in foundational models. Large model startups must also confront the challenge of whether they can persistently advance model iterations and explore viable application scenarios.

These factors depend heavily on whether the capital markets are willing to provide long-term, stable financial support.

In April of this year, KNOWLEDGE ATLAS initiated an A-share listing tutoring filing with the Beijing Regulatory Bureau of the China Securities Regulatory Commission. However, as of December 12, the company had not received any further feedback or inquiries from the CSRC regarding the advancement of its A-share listing.

In this context, KNOWLEDGE ATLAS opted to pivot to the Hong Kong market, seeking more sustainable fuel for this high-investment, long-cycle large model race. Simultaneously, it will face a dual test of fundraising capability and market confidence—whether anyone is willing to pay for the long-term investment in AI.

From GLM to MaaS: KNOWLEDGE ATLAS's Large Model Technological Foundation and Commercialization Path.

The prospectus indicates that KNOWLEDGE ATLAS primarily provides services ranging from computing power and API interfaces to MaaS (Model-as-a-Service), supporting both on-premise and cloud deployment models, and has been deployed across multiple industries.

As one of the representative companies in China engaged in the research and industrialization of general-purpose language models, KNOWLEDGE ATLAS's technology system is centered around GLM, covering text, multimodal, and application-oriented model services.

GLM belongs to the Transformer-based large language model paradigm, achieving unified modeling for both understanding and generation tasks by combining autoregressive generation with masked prediction. This architecture was initially proposed by KNOWLEDGE ATLAS in collaboration with related research teams from Tsinghua University and has been iterated upon in subsequent models.

In 2021, KNOWLEDGE ATLAS released China's first proprietary pre-trained large model framework, GLM, and launched a Model-as-a-Service (MaaS) product development and commercialization platform to provide large model capabilities and services externally.

In 2022, KNOWLEDGE ATLAS released and open-sourced GLM-130B (a bilingual Chinese-English model with hundreds of billions of parameters), marking the formal application of the GLM system to pre-trained large language models.

In January 2024, the GLM series reached a significant milestone with the launch of GLM-4, which supported longer context lengths, offered faster reasoning speeds, and substantially reduced inference costs.

In July 2025, KNOWLEDGE ATLAS further open-sourced GLM-4.5. Within 48 hours of its release, this model topped the global trending chart on Hugging Face, the world's largest open-source model platform.

In September of the same year, KNOWLEDGE ATLAS released and open-sourced GLM-4.6. As a further upgraded version of the foundational model, GLM-4.6 primarily enhanced coding capabilities. In November, GLM-4.6 ranked first globally on CodeArena.

In December, KNOWLEDGE ATLAS launched its latest flagship model, GLM-4.7:

In core coding capabilities, compared to the previous generation GLM-4.6, GLM-4.7 achieved significant improvements in multilingual agent programming and terminal-based tasks, with scores of SWE-bench 73.8% (+5.8%) and SWE-bench Multilingual 66.7% (+12.9%).

In ambient programming: GLM-4.7 achieved a major leap in UI generation quality, capable of producing cleaner, more modern web interfaces and providing more accurate layout and size control in presentation generation, resulting in an overall superior visual effect.

In tool calling: The tool usage capability of GLM-4.7 was significantly enhanced, demonstrating stronger practical operational abilities in web browsing tasks covered by the BrowseComp benchmark.

In complex reasoning: It scored 42.8% on the HLE (Humanity’s Last Exam) benchmark, an improvement of 12.4 percentage points over GLM-4.6.

Compared to models like GPT-5, GPT-5.1-High, Claude Sonnet 4.5, Gemini 3.0 Pro, DeepSeek-V3.2, and Kimi K2 Thinking, GLM-4.7 also demonstrated outstanding performance.

Simultaneously, KNOWLEDGE ATLAS also released multimodal models for different functions, including CogView (image generation), GLM-4.5V (visual understanding and reasoning), CogVideoX (video generation), and others.

In the AI Agent domain, KNOWLEDGE ATLAS's foundational agent model is AutoGLM. In December, KNOWLEDGE ATLAS fully open-sourced the core models of AutoGLM, marking a further development of AutoGLM within an open ecosystem.

As of June 30, 2025, KNOWLEDGE ATLAS models had provided support for over 8,000 institutional clients; as of the latest practicable date, they supported approximately 80 million devices.

On the commercialization front, KNOWLEDGE ATLAS began laying out its MaaS business model as early as 2021.

The MaaS platform primarily offers four categories of model capabilities, covering core areas such as language models, multimodal models, agent models, and code models, while also providing an integrated toolchain supporting model fine-tuning, model deployment, and agent development.

From the expansion of model capabilities and the advancement of agent technology to the gradual formation of the MaaS commercialization system, KNOWLEDGE ATLAS has completed a relatively comprehensive cycle of technological and product layout.

However, the finalization of the listing does not mean the game is over. As it steps into the public market, the intense R&D expenditures, escalating computing costs, and the reality that the commercialization path for general-purpose large models is not yet fully proven will all be subject to more transparent scrutiny.

Listing is not the finish line, but rather the beginning of a longer-cycle public test.