MININGLAMP-W (02718) announced its mid-year results for 2026: revenue from intelligent service integration grew by 605.7% year-on-year, and the FDE model is accelerating its advancement.

MININGLAMP-W (02718) officially released its interim results for the six months ending June 30, 2026. During the reporting period, the company achieved revenue of 760 million yuan, a year-on-year increase of 18.0%; gross profit was 390 million yuan, up 8.3%; and the net loss narrowed to 64 million yuan, a decrease of 68.5% year-on-year. During the reporting period, Minglue Technology's AI Native transformation accelerated further. Revenue from Agentic Services (intelligent agency services) saw substantial growth, becoming a key business segment driving the companys revenue increase; the open-source trusted Agent collaboration network Octo was officially launched, with a series of end-side models being deployed; the AI organizational restructuring and FDE layout proceeded in tandem, with a comprehensive strategic deployment for the new intelligent economic form fully underway. In the first half of the year, revenue from marketing intelligence services reached 344 million yuan, a year-on-year increase of 4.2%; the smart store operation system has covered over 75,000 offline stores. The long-accumulated data capabilities and industry experience have supported the entry of intelligent agency services into specific business segments. Multi-scenario deployment of intelligent agency services has become the main engine for revenue growth. During the reporting period, the revenue from the company's intelligent agency services reached 171 million yuan, a year-on-year increase of 605.7%, accounting for 22.5% of total revenue, becoming the main source of the companys revenue growth; new customer proportion reached 68%, with the customer base continuing to expand rapidly. The company applies its self-developed AI agents to strategy generation, content production, and execution, enhancing delivery efficiency while expanding the customer base, further validating the commercial feasibility of a performance-based payment model. In the marketing content production scenario, the automated production line can continuously output dozens of short videos that meet advertising standards in a single day, with the copy approval rate improving to 80%, helping brands increase content testing frequency and reduce production costs. New demands arising from AI search and content cognition are also being released at an accelerated pace, with relevant services covering over 100 leading clients across 15 industries. Influencer marketing business is gradually shifting from single project collaboration to platform-based operations. During the reporting period, the repurchase rate for related business clients reached 90%, with the average transaction value increasing by 2 to 3 times year-on-year. As centralized procurement increases on platforms like TikTok, influencer resources are transitioning from one-off project resources to sustainable organizations capable of matching and reusing content supplies. The AI-native foundation continues to be strengthened, with Octo promoting multi-Agent collaboration within organizational workflows. During the reporting period, Minglue Technology officially launched its first open-source trusted Agent collaboration network, Octo, aimed at building "the Internet of Agents" and upgrading AI agents from personal efficiency assistants to organizational-level digital labor forces that are collaborative, orchestrated, and accumulative. Octo provides agents with identity recognition, task orchestration, context sharing, and work record capabilities through product concepts and functional designs like Bot, Channel, Thread, and Loop, enabling specialized Bots to perform distributed collaboration under human guidance, and forming collective intelligence that exceeds the capabilities of any single model through network effects. Octo is open-sourced under the Apache License 2.0 and supports private deployment, coupled with a clear design for data ownership, allowing data sovereignty and knowledge entities to genuinely return to users and organizations, while also protecting, amplifying, and passing on individual tacit knowledge while respecting data boundaries. Octo efficiently connects personnel within enterprises with Agents, while Minglue Technologys self-developed first AI-native recording hardware, Octic, further integrates high-value context generated from offline meetings and real interactions into this collaboration network, forming a closed-loop and ecological synergy of "B-end data assets + C-end product experience." Currently, Octo has coverage across web, desktop, mobile (iOS and Android), browser plugins, and command-line interfaces (CLI), achieving multi-terminal interoperability. The Octo platform demonstrates stability in high-concurrency tasks and complex environments involving multiple members and Agents. Within the company, over 1,800 employees and 5,100 Agents are already engaging in seamless collaboration through Octo. Additionally, the company is closely engaging with major clients across various industries to promote the deployment of Octo in scenarios such as content production and organizational collaboration. In the future, Octo will continue to adopt an open-source approach and co-create new collaboration methods for AI-Native organizations with developers, corporate clients, and ecological partners, promoting the implementation of trusted, controllable, and sustainable privatized Agentic AI in real work scenarios. Technology-driven efforts strengthen the capability building of general models on the client-side. As large models penetrate deeply into enterprise scenarios, issues such as data security risks, ongoing inference costs, and response delays in cloud deployment models have become increasingly prominent, making edge deployment an important development direction in the industry. Based on an in-depth exploration of the MoE architecture inference mechanism, the company has proposed a new paradigm for Scaling Out model construction. Models built through Scaling Out perform on par with leading models in targeted capability domains, and significantly outperform distilled models of the same parameter scale, achieving a clever balance between model scale and capability precision. Using this method, Minglue Technology has developed several client-side general models covering multimodal understanding, including GUI visual-language-action, image-text and voice; Agent operations, including code generation, tool invocation, and contextual learning; as well as core capabilities for general cognitive reasoning. Relevant models can be customized to combine different capability domains according to clients' specific scenario needs. Furthermore, the company has developed and open-sourced the edge inference acceleration SDK Cider based on the Apple MLX ecosystem. This SDK is deeply optimized for macOS and Apple Silicon platforms, filling capability gaps in activation quantization and specific tensor computation within the native MLX framework. By improving edge inference efficiency, Cider lowers the deployment threshold for edge models, providing an engineering basis for their operation on personal computers and other edge devices. Continuously build AI Native organizations, FDE models accelerate on-site customer engagement. The company upgrades AI from a single-point tool to an underlying productivity that permeates the organization, advancing AI Native transformation along three main lines: "enhancing organizational efficiency, empowering services, and capitalizing capabilities." During the reporting period, the company established a quantitative baseline for AI efficiency improvements across 172 reported delivery scenarios, achieving a comprehensive efficiency improvement of 56.1% and accumulating 359 reusable Skills, transforming individual experience via Skills and Bots into organizational assets. Some business lines saw a 40% year-on-year reduction in human delivery costs while increasing service volume by 40%. For enterprise AI to truly enter business scenarios, it requires a deep understanding of customer needs, which may not yet be fully articulated, to organize industry experience, business processes, and data permissions, and then integrate Agents into actual workflows. During the reporting period, approximately 30% of the companys internal delivery personnel transitioned to front-end deployment engineers (FDEs). With the support of Octo, FDEs work alongside clients to clarify demands, translate business objectives into specific processes, and organize models, data, and different functional Agents to drive enterprise AI into customer business scenarios. In addition, the company intends to leverage Pansoft's (300996.SZ) existing product capabilities, client scenarios, and implementation experience with a dual-driven approach of "AI + Xinchuang," further deepening into core ERP and a broader range of business system operations in large enterprises, using full-scenario intelligent empowerment to expand the vast incremental space characterized by high stickiness and high customer transaction value. Simultaneously, the company will promote the transition of the Pansoft team towards the FDE model, enhancing product and service capabilities that are more closely aligned with customer business scenarios to assist large corporate clients in their intelligent transformation. Minglue Technology is continuously validating the industrial value of trusted Agentic AI through real business closed loops. Focused on the long-term goal of "creating organizational intelligence in the AI era," the company will continue delivering results in real tasks and expanding the scale of applications, allowing expert experience, production processes, and collaborative relationships to become organizational capabilities that can be continuously accumulated and utilized.