Dongwu Securities: Smart Spectrum (02513) From Tsinghua Labs to Hong Kong Stock AI Upstart Focus Model Iteration and Ecological Flywheel

The Zhitong Finance App learned that Dongwu Securities released a research report saying that it is optimistic about the advantages of Zhi Specu (02513) AI in terms of local big model technology, open source ecosystem layout, and the ability of government and enterprises to implement localization. The localization business has maintained steady growth, the cloud business has become the main driving force, and believes that the company is expected to benefit from the long-term trend of China's big model industry's transformation from localized deployment to cloud services. It is recommended to pay attention.

The main views of Dongwu Securities are as follows:

Company profile

Zhi Spectrum AI is a leading independent general model developer in China. It was founded in 2019 and transformed from the technical achievements of the Tsinghua University Knowledge Engineering Laboratory (KEG). The company developed its own GLM (GeneralLanguage Model) pre-training framework. Unlike mainstream GPT architectures, it uses an autoregressive fill-in-the-blank design, and has unique advantages in understanding long texts, logical reasoning, and low illusion rate. The company adheres to the parallel strategy of open source and commercialization, and has built a complete model matrix covering the fields of language, multimodality, code and intelligence. The flagship products GLM-4.5 and GLM-4.7 rank at the top of the open source model in many international benchmarks, and have been widely recognized by the global developer community.

According to Frost & Sullivan data, based on 2024 revenue, the company ranked first among independent GM big model developers in China and second among overall GM big model developers, with a market share of 6.6%. As of the first half of 2025, the company has served more than 8,000 institutional customers, and 9 of the top ten Internet companies in China use the GLM model. The open source model has been downloaded more than 45 million times worldwide, the MaaS platform has more than 2.7 million registered developers, and the average daily token consumption is growing rapidly. It reached 4.2 trillion dollars in November 2025, and paid API revenue exceeds that of all domestic models combined. The company passed the Hong Kong Stock Exchange hearing at the end of 2025 and will be listed on the Hong Kong Main Board on January 8, 2026, with the stock code 02513.HK, becoming the world's first major model company to be listed.

The business model is centered on the MaaS (Model as a Service) platform, two-wheel drive localization deployment and cloud deployment

Localized deployment targets government and enterprise customers, providing privatized operation and customized services. The customer unit price is high and gross margin is stable, accounting for 84.8% in the first half of 2025, with a gross profit margin of 59%. Cloud deployment services companies and developers through API calls and subscriptions. The access threshold is low and scalability is strong. In the first half of 2025, it accounted for 15.2%, and the revenue share increased rapidly. The company takes the cloud business as a long-term strategic priority and aims to increase the share of API revenue in the future. Open source strategies are effectively leveraged, and the recently launched Zcode (AI code editor) and Zread (code base analysis tool) further enhance developer stickiness and promote paid conversion. Historical financial performance is characterized by high revenue growth. Revenue for 2022 to 2024 was 57 million yuan, 125 million yuan, and 312 million yuan respectively, with a compound growth rate of over 130%. Revenue for the first half of 2025 was 191 million yuan, a year-on-year increase of 325%, surpassing the full year of 2023.

The issue price of the Hong Kong stock IPO was HK$116.20 per share, and 374.19,500 H shares were sold globally, raising a net capital of about HK$4.3 billion, with a market value of about HK$51.1 billion after raising capital. The capital raised is mainly used to enhance general model research and development (about 70%), optimize MaaS platform infrastructure (about 10%), expand ecological cooperation and strategic investment (about 10%), and supplement working capital. Cornerstone investors include institutions such as Shanghai Gao Yi, Guangfa Fund, and Taikang Life Insurance. The subscription amount accounts for about 70% of the sales volume. The shareholding structure is stable. The founding team controlled about 33% of the shares through a concerted action agreement, and well-known investors such as Meituan, Alibaba, Tencent, Xiaomi, and Gao Lin took the shares to provide scenario support and strategic collaboration for the company.

Outstanding core competitiveness is reflected in the full-stack self-developed technology system, leading model performance, open source ecosystem, and deep adaptation of domestic computing power

The company accounts for 74% of R&D personnel. The core team comes from the Tsinghua KEG Laboratory and has extensive academic experience in the field of natural language processing. The GLM series iterates rapidly. GLM 4.7 has outstanding performance in programming scenarios, uses an interlaced thinking mechanism, and is highly stable in real software engineering tasks. Autoglm enables AI to operate smartphones and computer GUIs autonomously, opening a new paradigm for agents. Multi-modal capabilities cover fields such as Wensheng maps, Wensheng videos, visual understanding, etc., and CogView-4 and CogVideoX are at the forefront of open source reviews. These technological advantages have been transformed into actual commercial implementation, and are widely used in technology, finance, government affairs, medical care, manufacturing and other industries.

Profit forecast and investment rating: The bank expects the company's revenue to be 790 million yuan (+151% year over year), 1.55 billion yuan (+97% year over year), and 3.22 billion yuan (+108% year over year) in 2025-2027, respectively. The revenue structure gradually shifts from localization to cloud dominance. The bank expects overall gross margin to reach 50% in 2025, pick up and stabilize at around 51% in 2026-2027, gradually improve cloud gross margin from a low level to 40%, and gradually clarify the profit path. Comparable company valuations show that Zhipu's 2026 PS is 30 times higher (using IPO pricing divided by expected revenue in 2026) is higher than that of Shangtang Technology and the fourth paradigm (using market value divided by 2026 Wind's consistent expected revenue, where Wind's consistent forecast is as of 2026/1/6), but as a pure model player, it benefits from cloud scale effects and agent/programming scenario dividends. As revenue grows rapidly, PS has more room for rapid compression below 30 times, and the valuation positioning is reasonable.

Risk warning: model iteration risk, price competition risk, government and enterprise customer demand fluctuation risk, customer concentration and contract renewal risk, overseas expansion risk.