Mininglamp Technology

was live ·

Minerva Technologies 2025 Annual Earnings Presentation

[AI Key Points Summary]

Financial Performance

Achieve the first adjusted scaled profitability in 2025, generating over 40 million in profits

Core strategic revenue grew by 8.5%, reaching 1.36 billion yuan

- Gross margin increased by 3.8%, with gross profit reaching 790 million yuan, up 10.8%

- Cash reserves of 1.38 billion yuan, indicating a healthy financial position

Business Progress

- Successfully transformed from data intelligence to agent-based artificial intelligence services

- The newly added agent-based service business achieved a revenue breakthrough of 100 million yuan

- Launched the Deep Miner product, significantly improving data analysis efficiency

- Maintained the top position in the OS World ranking for edge-side models

Guidance for next quarter’s performance

- The agent-based service business is expected to see substantial growth by 2026, becoming the primary revenue contributor

- The marketing segment within the data intelligence business remained stable

- Gross margin is expected to continue increasing

- Increased investment in R&D and computing power

Opportunity

The agentic AI track represents a market worth hundreds of trillions of US dollars, with the marketing segment valued at hundreds of billions

Developing edge-side models and multi-agent collaboration platforms

Collaborating with partners like Tencent Cloud on computing power deployment

Significantly reducing costs through AI-driven efficiency improvements, with administrative expenses cut by 32.4%

Risk

The software industry faces disruption from AI, with traditional software companies potentially facing existential crises

Systemic uncontrollable risks and data privacy issues brought by AI agent services

[AI Conference Record]

Operator

Dear shareholders, investors, analysts, and friends, good morning. Welcome to MINGUAL Technology's 2025 annual earnings release. First, allow me to express, on behalf of the company and management, our heartfelt gratitude for your long-term attention and support of MINGUAL Technology

The management attending today’s release includes Founder, Chairman, Chief Executive Officer, and Chief Technology Officer Mr. Wu Minghui, hello everyone; Co-founder, Executive Director, President, and Chief Financial Officer Mr. Jiang Ping, hello everyone; Board Secretary, General Assistant, and Investment and Financing Lead Mr. Fan Xin, good morning everyone

This earnings presentation will be divided into four core segments: operational review, financial performance, future outlook, and Q&A. We will now move into the operational review segment. First, let us invite Mr. Wu Minghui to provide a comprehensive review of the company’s overall operations in 2025. Please welcome Mr. Wu

Minghui Wu

Dear investors, shareholders, and analysts, good morning. I am Wu Minghui, Founder, CEO, and CTO of MINGUAL Technology. Today, I am deeply honored to present our 2025 performance results. This is our first annual earnings release since going public, and I am very excited to have this opportunity to share with you all

First, let me share with you the business review section. It’s fair to say that 2025 has been an incredibly fruitful year for us. We achieved a significant historical breakthrough this year—our first adjusted large-scale profitability milestone, bringing in over 40 million in profits.

Of course, my colleague Jiang Ping will share more details on the financial aspects later. For now, I’d like to highlight one core achievement: despite considerable uncertainties about the future of the software industry as a whole, we still managed to achieve a 96% renewal rate with our major clients.

As you can see from our overall business process, we have introduced a new category called 'agentic service.' This marks a successful transformation and breakthrough for our entire business, which I will discuss in more detail shortly.

Within our overall business segmentation, in addition to our previous 'data intelligence,' we’ve added 'agentic service.' Meanwhile, you can also see an 'other' category in our reports. In our prospectus, this 'other' category corresponds to our industry solutions.

This portion actually represents a strategically contracting part of our business. Excluding this segment, the overall annual growth of our core strategic revenue was 8.5%, and our gross margin also saw substantial improvement.

Following the tremendous financial success we achieved in 2025, my focus today is to report on our strategic upgrade—from data intelligence to agentic service, or what we call agent-based artificial intelligence services.

If you recall, our prospectus mentioned that we are the largest data intelligence application software company in China. However, on the day of our IPO, we clearly stated in our company positioning that we are the world's first publicly listed company specializing in agent-based AI.

What’s the difference between these two? As you know, data intelligence essentially refers to the fact that what we provide to clients is not just software but software integrated with data. Many of our clients rely on us to collect data across various scenarios such as advertising and marketing, offline service operations, and more. We help them leverage AI technology for data analysis and mining, ultimately delivering analytical reports or providing online SaaS dashboards.

But as you all know, AI has revolutionized every industry, and there is no doubt it will transform our industry as well. Today, Minglue has completely evolved into a new model centered around 'agentic service.' It’s also for this reason that, after confirmation by our auditors, we’ve introduced a new category in our financial reports called 'agentic services.'

Recently, you may have come across some industry news or expert opinions. For instance, Sequoia Capital published an article suggesting that many software companies in the future will operate under the guise of service-oriented businesses but are, in essence, still software companies. I would say that our 'agentic service' aligns with this vision and puts it into practice.

At the same time, this track is also extremely large. Jensen Huang, the founder of NVIDIA, once mentioned that in the future, artificial intelligence essentially has two major tracks: one is called agentic AI, and the other is physical AI. Fundamentally, agentic AI refers to providing digital labor, meaning we will use AI to enhance white-collar work. Therefore, this track represents a market worth hundreds of trillions of dollars, presenting us with an excellent blue ocean opportunity for expansion.

During the IPO process, we were very fortunate to have selected an excellent stock ticker symbol, 2718. Since my undergraduate degree was in mathematics, I've always had a particular fondness for math, which influenced our choice of this ticker symbol. Perhaps many friends are unfamiliar with the meaning behind our ticker symbol.

2718 represents the base of the natural logarithm, or approximately 2.71828. The small 'e' in natural logarithms stands for exponential growth, which reflects our belief that with the development of artificial intelligence, especially agentic AI, our company and even the entire AI industry will experience significant growth in the future.

The function where this natural logarithm resides is e to the power of x, also known as the exponential function. In mathematics, it has an interesting property: the derivative of this function equals itself. Though we're not here for a math lesson, this aligns with the ancient Chinese saying, 'A craftsman must sharpen his tools if he wishes to do his work well.'

What does this mean? You can see that our core strategy throughout 2025 at Minglue revolves around refining our own product, called Deep Miner. This Deep Miner is actually the central tool in our original data intelligence business. By sharpening our own blade, the sharper it gets, the faster our business will grow in the future.

Thus, during the significant transformation we completed in 2025, we first ensured stable business growth while maintaining strong profitability. More importantly, over the past year, we have been refining our own highly effective tool, which can serve both ourselves and our clients.

Of course, from the perspective of agentic services, the future lies in serving customers through service-based models. Our focus isn't on software sales but rather on providing direct client services based on performance outcomes.

Let me give you an example. This chart represents a report we previously provided to a client after the launch of a 3C product. Don’t underestimate such a report. Below it, you'll find a line graph, and at each peak and trough of the line, there are consumer insights or industry trend observations.

In the past, when we provided data intelligence services to clients, creating such a report required a senior data analyst to repeatedly write code and adjust the data within our database, typically taking one to two weeks to complete such a report.

But today, I’m thrilled to share that thanks to our refined tool developed over the past year—Deep Miner—we can now generate such a report, like the one on this page, with just a single prompt and a keystroke, completing it in minutes. The efficiency is incredibly high, and this is just one example.

Everyone can see this page again. In fact, within our products, especially in our past services, such as social listening business, we have a wide variety of consumer insights, industry insights, and competitive analysis research tasks that we provide to clients. And these research tasks are gradually being empowered by deep miner, transforming into a model where we can quickly use AI agents to provide services to clients.

Of course, this kind of service model may gradually evolve into an agentic service mode in the future. That means, in the future, when our clients raise a high-level demand for data analysis or insight, the entire service process will be carried out by agents. Our human experts will only need to provide professional opinions in between.

After discussing the agent transformation in our data intelligence, let me report on the new progress in our agentic service business. We have achieved revenue from a completely new category. This revenue comes from delivering services to clients rather than directly providing software as I mentioned earlier.

In this process, we first focused on the marketing industry. There is actually a lot of room for development in agentic services, as Jensen Huang said, essentially this track represents the productivity market for all white-collar digital labor globally. This market space is very large, and starting from zero in 2025, we achieved a breakthrough with 100 million in revenue.

Why did we focus on advertising and marketing first? It's because Minglue has been established for twenty years. During these two decades, we have accumulated 2,000 global brands, including some of the largest brands in China. In this process, we have also gathered rich experience and diverse data in areas like consumer insights, ad placement materials, creativity, and e-commerce conversion in the marketing industry.

At the same time, this market has enormous potential. As everyone knows, the advertising and marketing market in China is worth trillions. Of this, there is a market scale of hundreds of billions that can be operated by service providers. Therefore, the service market scale far exceeds the scale of selling software or tools. Hence, we not only have our advantages here but also face a very large potential market. We are very honored to tell you that we have made a breakthrough from zero to one.

Let me show you. Our first application scenario is brand advertisement production, similar to a previous product we developed called Add If. The Add If product helps clients evaluate the effectiveness of high-quality brand advertisements. Sometimes, this agent might just be like a lobster, right? They can form a closed-loop interaction with today’s brand advertisement creation agents.

Welcome to watch this live demonstration. This video will showcase how the commercial video judging system based on Mano and HMLLM completes the full process from generation to evaluation. First, users input commands on the online agent platform, then the Mano model automatically takes over the user interface. In the GUI environment, the C Dance 2.0 system simulates real user operations on webpages, progressively submitting video generation tasks until a rough cut of a TVC advertisement is generated.

Next, users can issue further instructions. The system uploads the video to the Edit platform through Scale capabilities, and the HMLLM model randomly intervenes to perform content reasoning and analysis on the rough cut video. The system will automatically locate local video files and verify user requirements. Once confirmed, the task is officially submitted. During execution, the system demonstrates adaptive capabilities, handling complex issues such as Scale invocation logic and network proxies. The entire process requires no manual intervention.

Users can check task progress in the conversation interface. After analysis is completed, the judging system returns results containing subjective understanding information combined with objective features like visuals and audio. Based on this, the system generates a complete analysis report. The report clearly identifies existing problems in the current video, which serve as crucial references for subsequent optimization.

After obtaining the analysis results, users can initiate further optimization commands. The system will utilize the AI-powered editing software to perform detailed processing on the video. At this point, the Melo model is activated again, progressively executing various operations in a real user environment, including key steps such as file selection, video upload, and clip filtering. It can also automatically handle detailed tasks like subtitle removal and renaming. All clicking, inputting, and waiting processes are independently completed by the model without human intervention.

Upon completion of the process, the system will control the interface to save the streamlined video to a specified local path, successfully concluding the optimization task. In the final stage, users can submit another evaluation request, aiming to subjectively assess the effect of the optimized video. The commercial video evaluation system then activates the model again to conduct inferential analysis on the new video.

After a brief wait, the system returns the evaluation results. Compared to the published video, the overall feedback is more positive. The system further compares the differences between the two results, showing changes across various metrics. Users can also access richer information on the platform, including attention heat maps, trend curves, and in-depth analysis reports. This content will help users comprehensively understand the video's effectiveness, providing clear data support for subsequent creative iterations and distribution strategy optimizations.

What I just demonstrated is a complete closed-loop automatic iteration from brand advertisement production to evaluation. This topic is also the focus of my doctoral thesis at Peking University. I recently passed my preliminary defense at school, and it’s a great pleasure to apply some of my academic research work to the enterprise's business operations, creating significant commercial value.

Of course, what we discussed earlier pertains to brand advertising. However, as everyone knows, the largest market space today lies in performance-based advertising. Now, let me share with you our fully automated AI workflow, specifically for performance-based advertising in the content e-commerce field, covering everything from video creation to final ad delivery.

The content marketing industry is undergoing an unprecedented revolution. First, blockbuster success has been overly reliant on experienced veterans, making it difficult to standardize and replicate successful formulas. Second, there is a production bottleneck where demand for distribution materials is exploding, yet ARPU per material is plummeting, creating increasing pressure on productivity. Third, asset loss occurs when materials are discarded after use, taking away core expertise as talent leaves, leaving no accumulated assets behind.

Zanmiao uses AI to redefine content methods by building a content production system that spans pre-, mid-, and post-campaign stages. To solve the first challenge—how to deconstruct a blockbuster—all you need to do is upload a trending industry video or simply input its link, allowing the system to automatically retrieve and analyze it. Previously, studying a viral video meant repeatedly watching it, relying on intuition, and manually taking notes, but still struggling to pinpoint why it resonated with viewers. Zanmiao simplifies this process significantly.

We have developed a video decoding and attribution analysis system. Our model automatically breaks down videos frame by frame, identifying elements such as character expressions, product details, and camera angles. We systematically organize all these elements into structured components, with each frame automatically tagged using our proprietary molecular-atomic labeling system. Examples include labels for usage demonstrations, before-and-after comparisons, pain-point scenarios, etc., along with AI-generated natural language descriptions, such as 'a close-up shot of the influencer washing their face with rich foam, looking delighted.'

This isn't merely about tagging; it transforms the essence of viral videos into structured data that can be broken down, analyzed, and reused. Searching for a specific scene no longer requires scrolling through frames one by one. Instead, use natural language queries—for example, entering 'washing face with rich foam,' and the system instantly locates all relevant storyboard segments. This means we can not only dissect viral elements but also reuse them in new videos. From now on, virality is no longer a mystery—it becomes a science that can be studied, learned, and replicated.

Based on our analysis of trending copywriting, if I want to create a new video, I don't need to start from scratch. Instead, I directly reuse the logical structure derived from previous breakdowns of viral content. Rather than editing a single full-length video, we establish a modular asset generation system. Storyboard uploads are categorized by tags, and these tags are assembled to facilitate reuse, replacing specific materials while retaining the logic of viral success.

Using the same logic, I can generate multiple versions with a single click. For example, if I set it to generate thirty variants, the system will combine different materials while keeping the core logic unchanged, but each video will use different scene segments. Look, in just a few seconds, thirty videos are generated, and for each video, the duration, number of materials used, and their combinations are clearly visible. This is the leap from manual workshops to industrialized production, allowing creators to focus on creativity and strategy while leaving execution and mass production to the system.

This is our third major capability: material review. We have our own rules for materials, such as incorrect logo placement, colors not meeting specifications, or prohibited words in the copy. Traditionally, this relied on manual checks one by one, but with our solution, we use AI for preliminary reviews. Once the final product is uploaded, the AI automatically detects issues based on the provided rules and generates an initial review report. After the model's preliminary approval, it can proceed to the manual review process.

After uploading the final product, the system automatically sends a review notification to the designated reviewers. The reviewers can view the finished product and add their approval comments. If there are any modification suggestions, they can annotate directly. If the material has no issues, after confirming its accuracy, they click 'approve,' and the video will automatically sync to the material library with an 'approved' tag.

Our solution is not just a tool; it’s a content engine in the era of large models, enabling your team to transition from randomness to inevitability, from fragmentation to procedural workflows, and from manual workshops to industrialization.

Alright, everyone can see that this is our content e-commerce platform, which is just one form of performance advertising. In fact, on other performance advertising platforms like Xiaohongshu, we will also roll out more agentic services for our clients in the future.

Next, I’ll present a completely new business category, which goes beyond marketing and enters the entertainment and cultural sector. As you all know, today in the entertainment industry, games and short-form content represent a huge market opportunity.

Initially, we provided agentic services for marketing and advertising to these clients. However, we later realized that ad creation is essentially the same as creating short-form content. As you can see today, the capabilities of our intelligent agents can now help clients automatically create new plot scripts.

As you can see, all we need to provide is a basic creative idea, and then our intelligent scriptwriter takes over. It can automatically design the plot, set up characters, and predefine the entire outline, storylines, and storyboard. Finally, we connect it to our tattoo image and tattoo video models to generate the final short-form content. This significantly reduces the overall production and development costs of the short-form series.

In the future, this process can be integrated into a seamless workflow, creating a closed-loop effect that allows continuous iteration. In the digital world, any system capable of providing strong feedback, combined with our agentic service capabilities, can self-iterate and self-optimize. As long as our tools are sharp enough, we will undoubtedly experience exponential growth in the future.

This is actually the generation of our scripts. Here, you can see how we leverage the latest capabilities of our system to produce AI-generated images, and eventually videos. You may want to understand where we truly create value in this process. Historically, we have collected massive amounts of user data and user preference data from the internet.

Actually, the models for tattoo videos and images we use today are not developed solely by us. Instead, we collaborate with many industry partners. However, the most important thing is that we understand what content consumers enjoy. We can actually use our AI to act as a screenwriter, generate prompts, and finally integrate them with the latest generation of tattoo image and video models to produce complete storylines and full self-produced series.

So let me recap the significant progress we made in our four major business segments of agentic services in 2025, ranging from regular performance-based ads to content e-commerce ads, our traditional strength in brand advertising, and finally our new venture in short drama production and placement. All these areas have seen tremendous growth, and 2025 really laid a foundation. I believe that in 2026 and beyond, we will see another massive leap forward.

Next, let me talk about a key technological breakthrough. In terms of tech advancements, we also made significant progress in 2025. Recently, AI, including tools like 'Lobster' which I believe many investors, analysts, and shareholders are already using, has gained widespread popularity.

Within the industry, there's been a renewed and deeper understanding of 'agentic' technologies, especially over the past two months. As the CEO of the company, who also serves as the CTO, I am personally working overtime every day to push forward the development of our next-generation products. In fact, during the entire Spring Festival, I didn't get much rest and was working with the team to keep things moving forward.

In this process, we identified a huge opportunity, which we call 'proactive agent,' or active proxy intelligence service. Many of our investor friends might already be familiar with various smart agents used in B2C or B2B scenarios, such as DouBao YuanBao, which are essentially intelligent agents.

However, you can see that a completely free intelligent agent won’t provide the strongest model or fully showcase its best capabilities. Why? Because it’s free, the better the service it provides, the greater its losses, which presents a business model issue.

Even paid services, such as subscription services like GPT at $200 per month, struggle to become proactive agents. What exactly is a proactive agent? For example, if I were your financial advisor or agent—a proactive agent—it would mean that even if my clients don’t request anything specific, I should still study new investment opportunities and potential portfolios for them 24/7.

But as you all know, this process consumes a lot of tokens. So if a client pays me a fixed amount, say $200 per month, and I actively provide services as a vendor, I would still be losing money. The more I do, the worse my bottom line becomes.

There’s another form of service on the market where services are provided based on usage charges or token consumption. This creates a problem: clients hesitate to use the service because each interaction consumes a large number of tokens, making the cost very high.

Based on this model, I've drawn a logical inference and prediction for the future. In the future, whether through edge computing models or hosting in the cloud—though the computing power would be exclusively mine—we’ll adopt what we call a PPC model. This will become the dominant model in the market. Essentially, each user, whether an individual or corporate user, will have exclusive access to a portion of computing power, and that computing power will belong to me.

After I have purchased all of this computing power, 100% of it, I can deploy my agents on it, allowing them to be truly proactive because the computing power is now owned by me. I expect it to work 24/7 and utilize all its capabilities fully.

We believe this is a future direction, and for this reason, the models built on this computing power should be deployed as small and medium-sized models on the edge. In this area, we have achieved excellent results over the past year.

Last year, we also released two edge models, Mano and Sito. These two models are well known, and the rankings they belong to, such as Mano in the OS World and Man to Web, especially the OS World ranking, which is currently one of the most prominent global benchmarks for computer usage.

Whether it's Anthropic or OpenAI, every time they release a new generation of their models, they announce to all their investors and users their ranking on the OS World list. I am very proud and pleased to inform our investors that our model consistently ranks first among small models, specifically edge-side models, on this list.

Even in the overall rankings, despite using a small model, we have continued to maintain a leading position. Therefore, there needs to be a renewed understanding of Minglue. We are not just an application-focused company; we are a company with strong foundational model-building capabilities. Going forward, we will have more developments in this field, and we look forward to our investors and shareholder analysts' anticipation.

That concludes my part. Thank you for Mr. Wu's sharing. Next, let's invite Mr. Jiang Ping to interpret the company's financial performance for 2025.

Jiang Ping

Hello everyone, I am Jiang Ping, CFO of Minglue Technology. I am here to report the overall financial situation of Minglue Technology for 2025 to our shareholders and investors. As Hui Ge mentioned earlier, our basic revenue has slightly increased by 3.2%.

However, the most important aspect is that our core strategic revenue grew by 8.5%. Our total core strategic revenue reached 13.6 billion in 2025. At the same time, we significantly reduced our other revenue streams, particularly from customized industry solutions, due to their lower gross margins.

Thus, our revenue from these areas dropped from 1.27 billion in 2024 to about 65 million in 2025. This growth aligns with what Hui Ge mentioned earlier: our primary growth driver comes from the expansion of agentic services, demonstrating Minglue Technology's successful transformation in 2025.

The success of this upgrade actually began with our investment in this area back in 2024, and by 2025 we had achieved the desired results. Additionally, in data intelligence, as everyone knows from our prospectus, we have two main divisions: marketing intelligence and operational intelligence. Marketing intelligence remains stable.

Operational intelligence has seen significant growth. The primary driver of this growth is from our drawing intelligence, which refers to the increase in our offline store operations, particularly related to work badge business. Another key development is that the company's gross profit margin has improved substantially.

Our total gross profit reached approximately 790 million, representing a 10.8% increase. The overall gross margin increased by 3.8%, nearly 4%. This improvement is largely due to the implementation of AI within the enterprise, especially through the deployment of agentic services across various business teams.

It’s almost an exaggeration to say that every employee now has access to powerful tools like deep money and other resources to enhance efficiency dramatically. Moreover, in 2025 we achieved profitability on an adjusted basis for the first time ever, demonstrating the company’s ability to make the organization healthier and sustainably profitable.

Let me share with you the level of our three major expenses. From the perspective of R&D costs, they continue to rise, mainly driven by the increased spending on cards and token fees. As mentioned earlier, the application of AI internally has significantly improved efficiency. Notably, our administrative expenses have been reduced by 32.4%.

In order to achieve even faster growth in 2026, we completed the recruitment of sales personnel in 2025. Our sales staff, from recruitment to training and transitioning into AI-native roles, are fully prepared. Our sales expenses grew by 37.5%.

Overall, our cash position remains very strong. It's worth noting that our operating cash flow turned positive for the first time in 2025, reaching 18 million in positive cash flow from operating activities. By the end of the year, we had reserves of 1.38 billion in cash.

The overall debt situation is also quite healthy. We have a manageable debt of around 250 million, making the entire company financially robust. That concludes my overview of our financial status. Now, let's invite Hui Ge to continue sharing.

Operator

Reflecting on the past helps us better prepare for the future. Next, let's welcome Mr. Wu Minghui to introduce the company’s strategic plans and directions for 2026. Please welcome Mr. Wu.

Minghui Wu

Dear investors, shareholders, and analysts, let me share some insights into our future prospects. 2025 laid a solid foundation for 2026, and we should see the growth flywheel take off. It’s like an exponential function. If you look at the curve of y = e^x, it is relatively flat on the left side, but once it crosses what we call the tipping point, growth accelerates rapidly, bringing entirely new momentum.

Let me share with you some of our major strategic plans. First, regarding agentic service, in 2025 we laid the most crucial foundation in the field of advertising and marketing, which is traditionally our strong suit. In terms of performance-based ads, brand ads, content e-commerce, and even slightly beyond the marketing track into the cultural and creative sector, specifically short dramas, we have successfully validated our model. Therefore, in 2026, our primary focus will be to scale up these achievements.

I believe these areas will gain momentum in 2020. More importantly, from the Spring Festival until now in 2026, we've continued to sharpen our tools, and we will soon release more significant products. These products will allow our company to expand our agentic service beyond the marketing domain. So, in the future, while today's agentic service at Minglue is mostly related to marketing, there will definitely be more directions ahead, potentially even expanding globally.

The second part relates to the strategy for our edge models, which I mentioned earlier. Edge modeling is actually a very, very significant strategy for us because, in my view, many AI services today are still not intelligent enough. Let me give you a simple example.

As Jiang Ping just mentioned, we have a product called 'Work Badge.' This Work Badge product can be used in offline stores or by many individual consumers. I'm currently holding one that we're testing. This product can record audio in various scenarios, similar to how Tencent Meetings or other online conferencing tools can record meetings.

However, when we transcribe these recordings into text, we often find numerous errors, especially with specialized terminology or familiar names. What is the core reason? The core reason is that the system doesn't understand the context or the specific scenario you’re in.

For instance, sometimes when discussing a business issue involving a client named PNG, the transcription might mistakenly interpret it as referring to a janitorial role. Such issues are quite common. Thus, a future trend would involve running such tasks on the device side.

Why run these tasks on the edge? Because on the device side, we can access each person’s unique data in their specific context. This data is highly sensitive and private, so ensuring privacy protection is critical. Data sovereignty and edge deployment go hand in hand.

We believe different enterprises and individuals will continue to deploy edge models. We will keep investing heavily in this area. This year, we will increase our R&D investment, including CAPEX, focusing on computational power deployment. We have many partners across the ecosystem, including cloud providers and our own shareholders, with Tencent being our largest shareholder. They provide significant support in computational resources through Tencent Cloud.

The third part I want to discuss is our open-source strategy, which ties into our ecosystem deployment. A prevailing concept in today's market suggests that 'the software industry is dead.' As someone who has worked in the software industry for twenty years, I agree with this broad perspective.

However, I’d like to report to our shareholders, investors, and analysts that Minglue is in a very advantageous position. Although historically we’ve developed much software, the core essence of what our customers purchase from us is data—data. Just like in the investment industry, consider Wind Info, Bloomberg overseas, and tools like Big Wisdom or Tonghuashun. Think about it—the software itself isn’t important; the key value lies in the data behind it.

Therefore, our original core business in data intelligence has been minimally impacted by this situation. So, what is our core strategy today? We are developing new software. If we want our clients to directly use our software, we will follow the model of open cloud and create truly large-scale open-source software, deploying these open-source solutions globally, which simultaneously expands our international operations.

On top of that, we will provide our clients with edge-side model services and, further, industry-specific intelligent agents or services like 'Vertical Industry Lobster.' This is actually a very significant strategic direction for us. Looking ahead, stay tuned as we will roll out more multi-agent collaboration platforms and additional open-source infrastructure projects on this scale, increasing our global influence.

Lastly, let me introduce our global expansion strategy. In fact, we have been exploring the globalization of our AI capabilities for many years. However, previously, due to various reasons—especially geopolitical factors and compliance issues like privacy protection—many Chinese companies faced significant difficulties in globalizing.

But after understanding our earlier strategies, you’ll see that our open-source approach offers one key advantage: global partners and clients can confidently use our software because it’s open source. For example, many startups in Silicon Valley are already using our large open-source models like Tongyi Qianwen and DeepSeek from China.

Moreover, many companies in China are also adopting open-source software such as Open Cloud. Open source transcends borders and represents true security and trustworthiness. At Minglue, we have always aimed to build trustworthy, data-driven artificial intelligence, which is central to our future strategic planning.

I am extremely pleased to report to our shareholders that we have made substantial progress in numerous areas, and we will gradually announce these developments. That concludes my presentation on our future outlook.

Operator

Through the above introduction, I believe everyone now has a deeper understanding of the company's operations and plans. Now, we will move into the Q&A session of this conference. Participants who wish to ask questions via phone should press the * key on their handset, followed by the number one. Online participants can submit text questions in the live interaction area or click the 'raise hand' button to request a voice question. Thank you.

Hello, if phone participants would like to ask a question, please press the * key on your handset first, then press the number one. Online participants can submit text questions in the live interaction area or click the 'raise hand' button to apply for a voice question. Thank you. Next, we invite the investor with the phone number ending in 6806 to ask a question. Please provide your name and organization first. You may speak now. Thank you.

Xiao Kai

Good day, Mr. Hui, Mr. Jiang, and Mr. Fan. This is Xiao Kai, an analyst from CICC. First, congratulations on the company's strong performance. We’ve seen impressive results from your new business initiatives. My question is about our agentic service business. As Mr. Hui just mentioned, we’re seeing the software industry shift from simply selling software to delivering outcomes.

Then I understand this is also the underlying logic of our agent business. My question is, under this paradigm shift, how do we evaluate Minglue's long-term potential market space and valuation system? Thank you.

Minghui Wu

Alright, this is a very, very important question, and it also involves the core strategic development direction of our entire company. As I mentioned earlier to our investor analysts and shareholders, in 2025, on our agentic service, it precisely verifies that we have adapted and followed the trend of our industry. We have even more forward-looking and started the layout of our own Deep Miner product in advance.

It is also because of this reason that our entire work is proceeding in an orderly manner. In terms of this track, it's a huge track. As Jensen Huang said, the future of artificial intelligence really has two major tracks: one is Agentic AI, and the other is Physical AI. The total addressable market for each of these tracks, globally speaking, is hundreds of trillions of US dollars.

If we only look at the marketing sector, there is actually a scale of hundreds of billions. It far exceeds the scale of selling software or tools. In this process, companies like Minglue have made significant advances in layout. More importantly, historically, we have accumulated a vast amount of data in the advertising field, and we have unique AI capabilities. For example, as everyone just saw, whether it’s my doctoral thesis or Miaoa, in this process, we are truly helping clients create agents.

When providing services based on agentic service, our agents need to operate various industry software. For instance, I need to use CapCut to edit advertisements, or Dreamlike to generate videos, etc. Many of the models within these, including the agent models, are self-developed by us, so we have a very strong competitive advantage in this part.

Our bold judgment is that today's AI, due to agent services—especially with the continuous progress of Lobster over the past two months—many advertising companies are increasingly utilizing such AI agent capabilities. Therefore, to some extent, the original service industry might be further squeezed by clients if everyone's ability to use AI is the same.

What does this mean? For example, if there are five advertising companies, and each company is merely using ordinary Lobster to provide advertising services to clients, then everyone may be doing so-called agentic service. However, since the capabilities of these five companies could be homogenized, clients might end up compressing your quotations because they know your cost structure has changed.

But what I want to say is that if a company like Minglue, which has its own unique data, self-developed models, and an exclusive multi-agent framework, can produce content quality far higher than other service providers. In this case, we might even be able to raise prices for our clients because what we deliver is much better than the best humans could previously achieve.

This is similar to what everyone knows in the reinforcement learning domain—AlphaGo beat Lee Sedol, AlphaZero subsequently defeated AlphaGo. In fact, many of our AI deployments in the future will follow this path—we aim to use AI to create agentic services that surpass human capabilities.

In this context, we believe that the future market space for our services might be even larger than the traditional market space provided by humans. That is my answer.

Xiao Kai

Alright, thank you, Brother Hui. I don't have any other questions on my end and look forward to more new developments from the company. Thank you, thank you.

Operator

Next, we invite the investor with the phone number ending in 0365 to ask a question. Please provide your name and organization first. Please go ahead, thank you.

Yang Lin

Hello, President Wu, hello, management team. I am Yang Lin, a computer analyst at Guotai Haitong. I have a question to ask, and I also thank the leadership for sharing the strategy just now. We want to analyze what the development expectations are for each of our core business segments by 2026. Additionally, how will AI substantially drive our overall revenue scale and gross profit margin? Thank you.

Jiang Ping

Alright, let me answer this question. Hello everyone, I am Jiang Ping. Just now, it was mentioned that this is a forecast of our revenue and gross profit as of 2026. Brother Hui has already discussed a lot, especially regarding agentic service. From the perspective of agentic service, it will significantly contribute to our revenue growth in the future. I won’t disclose specific figures, but it will definitely be a substantial increase. Therefore, the main revenue contribution in 2026 will come from agentic service.

We have the most resources in this area, as the first question also touched upon the market potential, which is very large. Our proprietary data, models, and multi-agent collaboration frameworks have been well developed. This has already been commercially validated in 2025, and by 2026, this segment will see further expansion.

Now, returning to our data intelligence perspective, our Deep Miner product serves as an engine. Whether it’s in data collection, data insights, or data generation, it provides significant empowerment to our data-driven marketing efforts.

Thus, the construction of marketing will continue to optimize and improve in terms of gross and net profit due to this empowerment. As for revenue, because we are shifting our focus from just providing insights to decision-making and action, revenue will remain stable.

As for operational intelligence, this segment will continue to grow. The smart badge business, which was mentioned earlier, holds another promising possibility—there is a high likelihood that our smart badge business will be promoted to the consumer market (to C).

Of course, the B2B segment has also grown significantly. The intelligent painting solutions for offline stores represent a very large business line, whether it's in the automotive, real estate, or financial sectors—it's a massive market across the board. Therefore, the intelligent painting business will see substantial and continued growth due to our core engine as well as our hardware products like employee badges, combining both software and hardware.

From the perspective of revenue growth, and then looking at gross margin, we expect the gross margin to continue improving. As mentioned earlier, during our marketing efforts, efficiency gains, particularly from our internal processes, will significantly enhance our gross margin.

At the same time, with regard to our agentic service, since 2025 is still an early stage for this sector, its current gross margin hasn't reached a high level yet. However, due to scaling effects, we expect that gross margin to rise further. In summary, thanks to the technological reserves built up over the past two decades, especially in such a vast marketing landscape, 2025 will see growth in both revenue and gross margin, with revenue growth accelerating faster.

Yang Lin

Okay, good, thank you, thank you.

Operator

Next, let’s invite the investor with phone number ending in 9995 to ask a question. Please provide your name and organization first. Go ahead, thank you.

Zhu Xueqi

Wen Li, hello, I am Zhu Xueqi, a computer analyst at CITIC Securities. First, I want to congratulate the company on its strong performance in 2025. It feels like the entire company is truly at the forefront of AI technology exploration. My question is this: we notice that overall R&D investment seems relatively controlled compared to peers. Yet, models like our Mono have achieved excellent results on various benchmark tests. Could management elaborate on how we maintain continuous technological leadership? Additionally, I observe that these models are relatively lightweight in terms of parameter count. Does this suggest that future focus may be more on edge-side applications?

However, we have also seen that models like our mono have achieved very good evaluation results on some assessment rankings. I would like to ask the company's management to elaborate on how we achieve continuous technological leadership. Additionally, I have noticed that the parameter count of these models is relatively small. Does the layout of our lightweight models indicate that we may focus on exploring some aspects on the edge in the future? That's my question, thank you.

Minghui Wu

Indeed, over the past few years, even before our IPO, the broader capital markets were not particularly friendly to Chinese tech companies. Our own cash flow, as Jiang Ping shared just now, has undergone significant positive changes. Operational cash flow has turned positive, and we are continuing to secure financing.

I want to say that indeed, in the past few years, especially before our listing, the entire capital market has been quite unfriendly towards Chinese technology companies. As for our own cash flow, particularly in the past year, as Jiang Ping just shared with everyone, we have seen a significant and positive change; our operating cash flow has turned positive, and we are also continuously financing.

In this process, we will definitely increase our investment in compute power. This year, we are certainly going to step up in that area. However, on the other hand, our company's R&D roadmap is going to be quite different from mainstream foundational large model companies in the market.

As everyone knows, these foundational large model companies generally need to reserve close to tens of billions in GPU resources, meaning their annual capex allocation could amount to several billion dollars to support such massive fixed asset investments required for training giant models.

Today, in the track where Minerva operates, we believe that every vertical industry should have its own unique model. Once you train this unique model, it can protect the true intellectual property of that vertical industry and prevent it from being taken over by foundational model companies.

We firmly believe that even within the same advertising industry, the logic behind ad placements on Xiaohongshu, Douyin, Kuaishou, video platforms, or WeChat Moments differs significantly. The content formats that consumers enjoy also vary, and ultimately the models they require are personalized. If we return to individual consumers, it’s even more evident — each person may need a different model.

So today, there's an emerging AI research paradigm in Silicon Valley called 'continuous learning,' accompanied by another important method known as 'next learning' or 'nested learning.' Currently, Minerva is actively researching the latest advancements in this field, adopting a very distinct roadmap for investment and R&D efforts.

Our approach begins with specific business scenarios that truly create value for investors. In fact, I'd like to pay tribute to two great companies and their leaders. One is Jensen Huang, who initially focused solely on gaming GPUs when Intel was making all-purpose chips. Today, NVIDIA has grown from that specific use case into one of the world’s highest-valued companies. I believe this is a viable roadmap.

The second example is DeepSeek. As many know, Liang Wenfeng’s team innovatively open-sourced the R-one model with only 2,000 training cards, which Minerva still possesses today, supported additionally by Tencent Cloud. Therefore, we strongly believe that great innovations don’t come merely from piling up resources. Instead, we have a clear and well-thought-out roadmap of our own.

Zhu Xueqi

Thank you, leader, for your response.

Operator

Next, let us invite the investor with the phone number ending in 7603 to ask a question. Please provide your name and organization first. Go ahead, thank you.

Yang Jiuyi

Alright, alright, everyone from Minglue, hello leaders. I am Yang Jiuyi, an analyst from the Huachuang Computer Group. I have a question that I would like to ask. As AI truly penetrates further into the core business processes of financial institutions and manufacturing clients, issues such as data privacy and system instability risks caused by multiple agents might become critical factors in client decision-making.

I would like to ask, what systematic arrangements has Minglue made in its underlying architecture to address and resolve these compliance-related concerns from our clients? I kindly request the leaders to answer. Thank you very much.

Minghui Wu

Alright, thank you. Let me continue answering this because it is related to our technical roadmap and product research. In fact, I have already emphasized one of our key directions: the open-source strategy. In my view, any software company that remains closed-source in the future, without its own unique data moat or model moat, will face a catastrophic disaster.

Actually, I have a friend who, during the Spring Festival, personally developed a CRM software through web coding. He is both the CEO and the CTO, as he is also the top technical leader in the company. He replaced their previous paid CRM system with this new one. So you can see that open source will certainly be a major trend in the entire software industry in the future.

From Minglue's perspective, we firmly embrace open source. Some of our smaller models will be open-sourced, and our larger software systems will gradually follow suit. Under the backdrop of open source, it will greatly satisfy the needs mentioned earlier, such as in finance and government sectors. More importantly, this allows us to embrace clients from all over the world.

Because, based on this, by open-sourcing and providing access to the source code, we can generate an extremely strong sense of trust. The transparency ensures everyone can see whether there are any issues with the code. However, as open source continues to develop, we still have many value-added services. These will form part of our continuous innovation in business models and revenue streams. Therefore, we solve compliance issues through open source.

On the other hand, Minglue places high importance on data privacy protection. Our ethics around artificial intelligence and data ethics were actually part of my first research topic when I was pursuing my Ph.D. at Peking University, where I studied AI ASICs, a field I highly value. You can see in our prospectus that there is a dedicated chapter on ESG aspects.

When most companies discuss ESG, they usually talk about carbon reduction and environmental protection, right? But the ESG section of Minglue’s report was personally written by me and is actually derived from part of my doctoral thesis research. It discusses how we can better benefit humanity in the age of artificial intelligence, giving humans better work experiences and improved quality of life while not losing our sense of existence or dignity, creating a beautiful world where humans and machines coexist. This is actually our long-term goal for the future.

Yang Jiuyi

Okay, thank you for the leader's response.

Operator

Next, let's invite the investor with the phone number ending in 7279 to ask a question. Please provide your name and organization first. Go ahead, thank you.

Chen Shiguo

Hello everyone, good morning to the management team. I am Chen Shiguo, an overseas analyst at TF Securities. First of all, congratulations on the company's achievements in the past 25 years. I would like to ask two questions. The broader context is that with the rise of Open Cloud, or Lobster, the industry is undergoing a technological reassessment. Previously, large models were mainly used for chatting, but now people are expecting them to actually accomplish tasks.

So my first question to the management team is: how do you assess the changes Lobster has brought to the industry and its underlying significance? Secondly, how will Minglue seize this opportunity and integrate various company strategies, such as serving B2B or B2C customers, to tangibly reduce costs and improve efficiency through its products and services? That’s all for my questions, thank you.

Minghui Wu

Alright, let me answer this. It’s a very hot topic right now because I’m also giving a lecture at a business school tomorrow morning, and many students are concerned about the future organizational structure and how it can be restructured given Lobster's powerful capabilities today. They’re also curious about my thoughts on the Lobster product.

Actually, I’ve already mentioned this earlier. Our most significant R&D focus throughout 2025 was on a product called Deep Miner, which is an agent. It is not a single agent but rather a multiple-agent system. If any of you have used Deep Miner, you’ll know that when you assign a task, there is a coordinating supervisor—an orchestrator—that breaks down complex tasks into subtasks and then calls on different agents to execute those tasks.

Currently, products like Deep Miner still operate within a cloud-based computer framework, and their permissions are limited. However, the Lobster product from Open Cloud grants agents much broader permissions—full computer usage, or what we call 'computer use.' It provides an agent with complete control over a computer.

This level of permission is very extensive, which leads to much greater capabilities, although it does bring some security concerns. So if you look back at our previous models, such as the Mano Mano model that topped the OS World list—a ranking focused on computer use—you’ll see what I mean.

Just the other day, Anthropic released a new Lobster-style product, whose core technology relies on a multi-modal VLA model like Mano to automatically operate computer systems. Earlier, I demonstrated how we utilize these capabilities in our Agentic Service, which leverages our self-developed models specifically tailored for the Chinese market. These models excel in navigating software designed for the domestic market.

As for Open Cloud, from a technical perspective today, it has to some extent achieved AGI for humanity, which is an important point of view that I hold. The foundational large models I mentioned earlier actually lack the ability for continuous learning; that is, once they are launched, they lose the ability to learn knowledge after their release. However, many researchers in Silicon Valley are now attempting to use new nested learning methods or our continuous approaches to enable these models to keep learning.

Every day, people give these models new prompts and requirements, allowing them to continuously iterate. Humans learn every day, but the foundational large models, whether from OpenAI or Anthropic, do not have this capability. However, if even the researchers in Silicon Valley achieve this ability today, it would actually be less friendly to humanity.

Why? Because our privacy would be completely compromised. Every day, all the prompts, words, content, information, and knowledge we provide to these large models are taken by them for continuous learning, eventually harvesting everything from humanity, which is detrimental to us.

Therefore, I think Open Cloud is a great product and discovery. It operates on the client side, enabling each individual user to store their daily knowledge and accumulation in plaintext memory on this computer as they use it. This computer understands you better and better, becoming increasingly familiar with you, and performs your tasks more efficiently over time.

However, the biggest issue today is that Open Cloud is essentially a personal assistant—a personal aide. Peter originally created this product for his own use, not as a tool for organizational participation. On the other hand, Minglue’s primary business is B2B, serving enterprises.

Thus, in the future, we will upgrade our Open Cloud-like open-source product in the direction of enterprise needs. Additionally, we will make significant breakthroughs in products within the multi-agent framework. We also look forward to our investors' anticipation of our upcoming releases of open-source products.

So, first of all, I highly recognize the Open Cloud product. Like Deep C1, it represents a great contribution to all of humanity. In the future, we will also contribute further iterations of our product, combined with Open Cloud, as open-source projects to society.

Chen Shiguo

Alright, thank you, leader, for your response.

Operator

Due to time constraints, let's take one last question. This is a text-based question from the web. The question reads: 'Dear management team, the company has recently made positive progress in its business, and the capital market is also very concerned about the company's liquidity situation. Could you please share the current progress and expectations regarding the inclusion of the company in Stock Connect? Thank you for your response.'

Fan Xinyu

Let me address and introduce this matter. Good morning, investors. Thank you all for your concern. Everyone can actually look at our stock code. There's a 'W' at the end, which means that the company’s structure is different from most companies listed in Hong Kong. We have WVR, or Weighted Voting Rights structure. Under this structure, the exchange has a set of relatively strict evaluation criteria for companies like ours. Specifically, after our listing, there will be an evaluation period of 183 days plus twenty trading days.

During this evaluation period, the company needs to meet the average daily market capitalization and overall trading volume requirements. After meeting these two key criteria, the exchange will issue an announcement regarding whether the company is included in the Stock Connect program.

Therefore, although the company has been included in the Hang Seng Tech Index, the timetable for inclusion in the Stock Connect has not yet arrived. Based on our November listing schedule and the timeline set by the exchange, theoretically, the company could be eligible for inclusion around early June. However, we kindly ask all investors and friends in the capital markets to follow our official account, especially the exchange’s announcements, for the final decision on inclusion. Thank you for your attention, and thank you.

Operator

Thank you all for participating, and thank you to the management for answering questions. Today's online press conference has officially concluded. If you have more questions, please feel free to contact the company’s investor relations department after the event. Thank you all, and thank you.

More details:MININGLAMP-W IR

Note: The above content is generated by an AI language model based on publicly available data and third-party automatic subtitles. The content does not represent any position of Futu, nor does it constitute any investment advice. Futu Group makes no express or implied guarantees or statements regarding the accuracy, timeliness, or completeness of the above content.

17K Views