Releasing AI PC Edge-Side Model; Mininglamp Technology (2718.HK) Makes Another "Strategic Move" on the AI Application Front

Recently, Mininglamp Technology (2718.HK) officially released the Mano-CUA-2.0 version, an edge-side GUI VLA model tailored for AI PCs, and concurrently launched its "Thinking Mode." The newly upgraded Mano-CUA-4B-2.0 is already available for download and deployment on Hugging Face and ModelScope.

Model Upgrade: From "Executing upon a Single Glance" to "Thinking Things Through Before Acting"

It is understood that while the industry generally defines an AI PC as a "personal computer equipped with local AI acceleration chips," Mininglamp Technology (2718.HK) proposed a different perspective: a true AI PC should not merely be a "computer with AI," but rather a "computer built for AI."

The core upgrade of Mano-CUA-2.0 is the debut of the Thinking Mode. Put simply, it enables the AI to transition from "executing actions the moment it glances at the screen" to "thinking through multiple steps clearly before acting."

On a technical level, Mano-CUA is a GUI VLA model possessing capabilities in perception, understanding, planning, operating, and verifying graphical interfaces. It directly controls desktop software, webpage interfaces, and complex workflows via a vision-only approach. In a test involving 100 real-device macOS GUI tasks (conducted on a MacBook Pro with an Apple M5 chip and 16GB of RAM), the Thinking Mode boosted the overall task success rate by approximately 9% compared to the Fast Mode; in medium-to-high difficulty tasks, the improvement reached 10% to 13%.

Crucially, the entire process is completed locally on the device, with no need to upload data to the cloud. This means Mininglamp Technology (2718.HK) has simultaneously crossed the two major hurdles for enterprises utilizing AI to replace manual labor: the reliability of complex tasks (whether they dare to hand them over) and data security compliance (whether they are permitted to use them).

This also implies that under usage environments such as network disconnections, local deployments, or privacy-sensitive conditions, edge-side devices can comfortably assume more complex business workflows, truly expanding the application potential of edge-side agents in scenarios like office automation, software operations, and privatized business workflows.

Targeting the Rigid Demand of "Data Remaining On-Premises" in Heavily Regulated Industries; The Global AI PC Market Penetrates Rapidly

In industries such as finance, government affairs, and healthcare, the implementation of AI consistently faces a practical question: can data leave the local environment? Banks have massive amounts of cross-system data entry, verification, and report generation work daily; hospitals need to handle repetitive workflows like medical record entry and medical insurance checking; and government systems often face issues of internal-external network isolation and numerous legacy systems. These tasks are naturally suited for automation, yet they are often difficult to connect directly to cloud-based large models.

The core value of Mano-CUA-2.0 lies in three aspects:

First, Compatibility. Adopting a vision-only GUI control approach, it can directly comprehend desktop software and webpage interfaces. When facing a large number of legacy business systems without standard APIs, the Agent can look at the screen and operate just like a human, saving the cost of interface adaptation.

Second, Stability. The Thinking Mode adds task planning and reasoning steps prior to execution, solving the issue where edge-side models were previously prone to missing steps or suffering execution deviations during long-process tasks.

Third, Cost Advantage. In cloud-based models, a thinking mode implies a lengthy chain of thought and massive token consumption. In contrast, edge-side computing power naturally operates on a buyout model—once the hardware assets are purchased, the extension of thinking depth will not incur any additional marginal costs.

IDC forecasts that by 2028, over 70% of enterprises globally will adopt hybrid or local AI deployment architectures to satisfy data governance and compliance requirements. The release of Mano-CUA-2.0 lands precisely at the node of this structural trend.

As the latest milestone in Mininglamp Technology's (2718.HK) edge-side model lineup, the open-source project Mano-P has been upgraded into an exclusive sequence dedicated to edge-side layouts within the Mano series. Moving forward, more Mano-P series edge-side models will continue to be released, driving edge-side AI to transition from single-model capabilities toward multi-model matrix collaboration.

The market foundation for this layout is shaping up rapidly. According to Gartner data, global AI PC shipments reached 77.8 million units in 2025, capturing a penetration rate of 31%; by the end of 2026, global AI PC shipments are expected to rise to 143 million units, with the penetration rate climbing to 55%. Data from LeadLeo Research Institute shows that the scale of China's edge-side AI industry was under RMB 200 billion in 2023, and is projected to break through RMB 1.9 trillion by 2028, representing a compound annual growth rate (CAGR) of 58%.

On June 10, the Ministry of Industry and Information Technology (MIIT) issued the Implementation Opinions on the Innovative Development of "Artificial Intelligence + Information Communications" (2026-2028), explicitly proposing to "vigorously develop products such as artificial intelligence smartphones and computers, smart home devices, and intelligent wearable devices," and requiring the formation of more than 30 high-value typical scenarios by 2028. The dual resonance of policy and market is creating unprecedented incremental space for edge-side models.

According to public records, through successive strategic deployments since 2026, Mininglamp Technology (2718.HK) has become the only AI company in the Hong Kong stock market to concurrently possess a four-tier capability spanning edge-side models, inference engines, Agent collaboration platforms, and hardware entry points. From open-sourcing Mano-P and Cider in May to releasing Mano-CUA-2.0 and launching the Thinking Mode in June, Mininglamp Technology (2718.HK) has completed two major iterations of its edge-side models within a short span of two months.

Now, as Mininglamp Technology (2718.HK) progressively rolls out more implemented application models, the company has completed the upgrade of its technical system from a "single model" to a "multi-model matrix" for edge-side AI. It also continues to validate the value loop from "technical narrative" to "commercial implementation." At the industrial turning point where edge-side AI moves from a "concept" toward a "trillion-RMB industry blueprint," Mininglamp Technology (2718.HK) has added another vital piece to its board.

Bohai Securities stated that with the continuous popularity of AI Agent applications, token invocation volumes have embraced substantial growth. As the capabilities of domestic large models continue to improve, the pace of their implementation across various application scenarios is expected to accelerate further, driving the commercialization process of the artificial intelligence industry forward.

Risk Warning: The above content is an objective compilation of recent public market information for reference purposes only and does not constitute any investment advice. Investing involves risks; caution is required when entering the market.

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.