The Trillion-RMB Blueprint of Edge-Side AI; Corporations Such as Mininglamp Technology (2718.HK) Deploy All-Round Multi-Dimensional Layouts

In the current development process of large models, edge-side AI is emerging as an important commercial evolution path. It shifts AI from its dependence on cloud servers to running locally on devices such as smartphones and PCs, offering faster response speeds, enhanced privacy protection, and lower ongoing operational costs.

This technical trend is resonating with the explosion of GenAI, driving edge-side AI from a concept into widespread deployment. Large model enterprises play a critical role here, connecting vast numbers of developers with downstream hardware manufacturers through methods such as open-sourcing models and opening platforms, to jointly construct a flourishing edge-side AI ecosystem.

Multiple Forces, Including Large Model Upgrades, Propel the Trillion-RMB Blueprint of Edge-Side AI

Presently, the market narrative of edge-side AI no longer remains in the speculation stage.

Recently, the latest data released by LeadLeo Research Institute shows that the scale of China's edge-side AI industry was under RMB 200 billion in 2023, but it is expected to surpass RMB 1.9 trillion by 2028, exhibiting a staggering compound annual growth rate (CAGR) of 58% from 2023 to 2028.

The core forces driving this growth stem from multiple main lines: first, definitive policy dividends; second, the continuous expansion of terminal product categories; and third, the upgrade of products and models.

In 2025, domestic authorities explicitly set quantitative targets in the "Artificial Intelligence+" Action Opinions, planning for the adoption rate of next-generation intelligent terminals to exceed 70% by 2027, injecting powerful momentum into market growth.

China Securities had pointed out that the market scale of edge-side AI will leap from RMB 321.9 billion in 2025 to RMB 1.22 trillion in 2029, with a CAGR of 40%. Policies also provide specialized subsidies for innovative hardware such as AI glasses, and IDC predicts that the shipment volume of AI glasses in the Chinese market alone will grow by 68% in 2026.

Calculated by the adoption rate, the scale of edge-side AI devices will reach 1.05 billion units in 2027 alone, corresponding to a hardware market exceeding RMB 500 billion. Morgan Stanley also pointed out that with the popularization of core terminals such as AI PCs (with a penetration rate potentially reaching 62% in 2026) and AI smartphones, superimposed by the explosion of AI wearables, the upgrade of the entire edge-side industrial chain will be immensely propelled.

Furthermore, along with the AI-driven transformation of operating system-level entry points such as Apple's WWDC, Apple's M5 series integrates a neural accelerator into every single GPU core, causing its AI performance to skyrocket by 3.5 times compared to the previous generation. Mac desktops with a unified memory architecture have become the most cost-effective choice for running local AI models, providing the foundational hardware infrastructure for the large-scale deployment of edge-side AI.

Among domestic large model enterprises, various "lightweight" models have also been launched to provide "brains" for edge-side devices. Among them, Tencent (0700.HK) Hunyuan introduced the HY-1.8B-2Bit "ultra-small" model, which has an equivalent parameter size of 0.3B and a memory footprint of only 600MB, boosting generation speeds by 2 to 3 times, and can be deployed directly on mobile devices.

Zhipu (2513.HK) has constructed a complete matrix ranging from edge-side small models to 100-billion-parameter flagship models. Its edge-side Agent, AutoGLM, has achieved autonomous execution across applications, and the company has partnered with automakers to build a domestic edge-side multimodal large model, applied in the rear-seat interaction systems of S-class flagship sedans.

In 2026, Mininglamp Technology (2718.HK) open-sourced its GUI agent model, Mano-P, which can run locally on Mac and ranks first globally on the OSWorld benchmark. It also launched Cider, an inference acceleration framework specially optimized for Apple Silicon, boosting operator speeds by 1.4 to 1.9 times.

Tech Giants Assemble, Seizing Strategic Positions Across the Full Chain from Cloud to Terminal

Since the beginning of 2026, the dense deployment by domestic tech giants in the edge-side AI field can be described as a strategic race.

Zhipu's AutoClaw is China's first "one-click installation" local version of OpenClaw, marking the transition of edge-side Agents from cloud inference to local execution.

Following the launch of its MaxClaw cloud AI assistant in February, MiniMax deeply encapsulated the open platform interfaces of its MiniMax Speech voice model and Music music model in March, further enriching edge-side application scenarios.

Cloud service providers such as Tencent Cloud, Alibaba Cloud, and Volcengine simultaneously launched OpenClaw deployment services on their cloud platforms, and accelerated the embedding of Agent capabilities across core entry points like WeChat and Taobao. ByteDance, on the other hand, is unleashing joint efforts through edge-side large models and AI hardware matrices, having already positioned itself in categories such as AI earphones and AI toys.

In this edge-side race, the strategic positioning of Mininglamp Technology (2718.HK) appears particularly distinctive. As a company that landed on the HKEX under the identity of the "first Agentic AI stock," it has currently erected a comprehensive product matrix on the edge side encompassing open-source models, inference frameworks, and AI hardware.

It is understood that its open-sourced edge-side GUI-VLA agent model, Mano-P, supports running locally on Mac, comprehending and operating graphical user interfaces via a pure vision approach. Ranking first globally on the OSWorld benchmark, it enables enterprises to achieve AI automation processing without uploading sensitive data to the cloud. Concurrently open-sourced was the Cider inference acceleration SDK, which releases the full computing power potential of edge-side INT8 hardware from the foundational layer, forming a significant differentiation barrier in Mac edge-side implementations.

TF Securities pointed out that according to Mininglamp Technology's 2025 performance announcement documents, both the company's Cito and Mano models rank among the top on international leaderboards. The company's business architecture is jointly composed of one technical foundation and two major business engines. The technical foundation, DeepMiner, is an enterprise-grade trustworthy AI agent platform independently developed by the company, and the business engines include data intelligence and agentic services. The core technical architecture of DeepMiner includes one hub and two major models, consisting of three parts: the first part is the multi-agent collaboration hub, Foundation Agent; the second part is the expert brain, Cito, focusing on the deep application of industry knowledge and professional reasoning; and the third part is the dexterous hand, Mano, focusing on the "last mile" of system execution.

On May 13, 2026, Mininglamp Technology officially launched Octic, its first AI Native hardware tailored for meeting scenarios, seamlessly integrating high-precision real-time voice recognition and in-meeting AI assistance features into its self-developed multi-Agent collaboration platform, Octo. This also signifies that the company has completed its three-tier product closed-loop from model to platform and then to hardware.

Looking at it current status, within Mininglamp Technology's edge-side AI product matrix, its Mano-P is responsible for the edge-side agent's understanding and decision-making, Cider is responsible for inference acceleration to release hardware computing power, Octo is responsible for multi-Agent collaboration and organization-level orchestration, and Octic is responsible for the data collection entry point of offline scenarios. At present, all four products across the model, engine, collaboration network, and hardware have been either fully open-sourced or officially released.

Jensen Huang redefined AI PCs at RTX Spark, but the infrastructure of edge-side AI does not merely comprise chips and models; it also includes how agents collaborate with each other, and how data from offline scenarios is collected and understood. Mininglamp Technology's layout precisely targets these core segments "outside of chips."

Large model enterprises are no longer satisfied with merely providing the "brain"; instead, they are deploying comprehensive layouts across multiple dimensions—including models, platforms, hardware, and application ecosystems—to deeply participate in and lead this monumental transformation of edge-side AI.

Bank of China International (BOCI) Securities expressed that artificial intelligence has officially entered the era of "Physical AI," and robots, autonomous driving, and edge-side AI hardware will become the core vehicles for AI.

NewTimeSpace Disclaimer: All content herein is the original work of NewTimeSpace. Any reproduction, reprinting, or use of this content in any other manner must clearly indicate the source as "NewTimeSpace". NewTimeSpace and its authorized third-party information providers strive to ensure the accuracy and reliability of the data, but do not guarantee the absolute correctness thereof. This content is for reference only and does not constitute any investment advice. All transaction risks shall be borne by the user.