Matters:

On February 7, 2025, Loongson Zhongke announced the launch of a device equipped with Loongson No. 3 CPU, which successfully started running the DeepSeek-R1 7B model. On February 23, Loongson Zhongke released a software and hardware full-stack inference all-in-one machine based on the DeepSeek large model, and the product realizes full-stack localization support from chips, systems to frameworks. At the same time, the "Beijing 2025 Terminal Equipment Centralized Procurement Project Demand Announcement Announcement" was released, and a total of 21,180 sets of Xinchuang terminal equipment were purchased this time, with Loongson loongArch accounting for 63%.

Comments:

Large model: full-stack independent technology, collaborative breakthroughs between domestic chips and AI large models. Based on the 3A6000/3C6000 series processors of the independent LoongArch architecture, Loongson Zhongke has successfully realized the localized deployment of the DeepSeek R1 7B large model, and completed the adaptation with the Taichu Yuanqi T100 accelerator card in only 2 hours, providing a faster, stronger and more economical training and inference experience. At the same time, ArcherMind Xinchuang computer and Wanglong computer with Loongson 3A6000 processor have realized the local deployment of DeepSeek, and there is no need to rely on cloud servers, avoiding service interruption caused by network fluctuations or server overload, and can efficiently complete document processing, data analysis, content creation and other tasks, significantly improving work efficiency. At present, substantial progress has been made in the collaborative adaptation of domestic chips and AI large models, laying the foundation for building an independent and controllable AI technology ecosystem.

All-in-one machine: Loongson's DeepSeek inference all-in-one machine was released, and the deployment of AI in government and enterprises has entered a new stage of security and controllability.

The Loongson DeepSeek software and hardware full-stack inference all-in-one machine can be equipped with 2 Loongson 3C5000 processors and supports up to 4 Taichu Yuanqi T100 accelerator cards; It supports all DeepSeek models (7B/8B/14B/32B/70B), and users can flexibly adjust the configuration of the all-in-one machine according to actual application scenarios and performance requirements. Through self-development and the introduction of third parties, it supports the implementation of common AI agent applications, such as code writing plug-ins, document generation tools, meeting quality inspection, intelligent government affairs, intelligent approval, financial data analysis, compliance review, enterprise customer service, office assistants, enterprise internal knowledge consulting, and other AI agents, to meet the needs of business scenarios in various fields such as education, government affairs, enterprises, and finance. Loongson Zhongke is working with partners such as Taichu Yuanqi, Cambrian, Tiantian Zhixin, Computing Energy Technology, openEuler and other partners to realize full-stack localization support from chips, systems to frameworks, completely get rid of dependence on foreign technologies, and enter a new stage of security and controllability in government and enterprise AI deployment.

Xinchuang: The policy is strongly supported, and the large-scale volume of domestic CPUs can be expected. According to the release of the results of Beijing's 2025 Xinchuang terminal volume procurement, Loongson won the bid for 13,336 devices (including 11,787 desktops and 1,549 notebooks) with a 63% share, far exceeding the 37% share of the ARM technology route. The Beijing centralized procurement covers 45 directly affiliated units, including civil affairs bureaus, social welfare hospitals and other livelihood institutions, forming a benchmark demonstration. Combined with the recent party and government orders of more than 45,000 units in Yunnan and more than 50,000 units in Jiangxi, the company's terminal shipments are expected to achieve a breakthrough in 2025, accelerating the iteration of domestic CPUs from "usable" to "easy to use".

Investment suggestion: Taking into account macroeconomic changes and the company's performance expectations, we adjust the company's profit forecast, and expect revenue in 2024-2026 to be 5.06, 8.07, and 1.090 billion yuan (7.14, 10.02, and 1.384 billion yuan before 2024-2026), a year-on-year increase of 0.1%, 59.3%, and 35.1%; It is estimated that the net profit attributable to the parent company in 2024-2026 will be -6.19, 0.12, and 98 million yuan (-0.96, 0.02, and 81 million yuan in 2024-2026), a year-on-year increase of -87.9%, 101.9%, and 712.7%, maintaining a "strong push" rating.

Risk warning: market development is less than expected; intensified competition in the market; The customer concentration is high.