Meitu Inc Aktie

Good afternoon. Welcome to the Meitu 2025 Annual Results Presentation. We provide English simultaneous interpretation. If you'd like to listen to the interpretation, please log in through Zoom and click the Interpretation icon, select English channel. If you do not click on this channel, you will not hear the simultaneous interpretation.

Today, the leaders participating include Mr. Xinhong Wu, Founder, Chairman and CEO of Meitu.

Greetings.

CFO and Company Secretary of Meitu, Mr. Gary Ngan; CPO and President of the Imaging Products Division, Mr. Chen Jianyi.

A warm reminder, we include predictive statements, and many risk factors that we may not control may influence these predictive statements.

Online participants, you need to rename and add your real name and affiliation name. Before giving your question during the Q&A session, please first clarify your affiliation and your real name. You may also provide your questions through audio or text in the comments section.

Firstly, we provide information about our company.

Investors, analysts and audiences, thank you for your participation. Before we formally start, I would like to extend about the announcement yesterday on behalf of the company. We found that some part of the results were published on social media. We investigated immediately after ensuring we found there may be some misoperations and functions in the process and some information was published on the account, and the responsible person deleted immediately.

We took two effective matters. First, we deleted all content on the platform, and we applied for the suspension in the Hong Kong Stock Exchange. What's more? Relative information was just shortly spread on March 25th, and we ensure the smooth function on 26th. We feel sorry for all the inconveniences caused to our investors.

Thank you all for joining today's earnings conference. Over the past few months, as generative AI technology has been evolving rapidly, there are various debates around the future of the AI industry. As you may have seen in the opening video through the QR code of -- that made the fastest moving trains, for example, from single model approach to model-agnostic strategy from AI-based apps to AI agents, from subscription to consuming tokens, from the form of our team, from mature teams to AI-driven teams.

We reached a strong momentum. The adjusted net profit attributable to owners of the company grew 64.7% year-over-year to RMB 907 million (sic) [ RMB 970 million ] exceeding our latest thoughts on the AI applications. First, Meitu is an AI application company focused on the photo and video industry. Based on AI capabilities and efficiency-driven organization, we focus on high-value verticals and build industry-specific AI agent teams to help address users' end-to-end content creation needs, delivering high-quality output.

In each specific use case, our AI agent teams are composed of multiple agents with distinct roles. They are responsible for efficiently executing different standardized repetitive tasks. Users can focus on creative ideation and decision making. The core advantage of the AI agent teams lies in their clear division of responsibilities and targeted training. Break down into multiple stages and assign corresponding roles and responsibilities to each agent, so we can ensure the teams perform better.

Besides combining years of user insights, accumulated industry know-how and the ability to flexibly deploy models, we constantly translate AI technology into reliable product capability. Additionally, different models have different strengths. There is not a model that can cover all the areas suited for e-commerce styles. Vertical companies combining different models, it will be more common for companies to deploy models. We will focus on more flexible and global things. Centered around it, we will build AI teams meeting industry needs.

Now, let me invite our Chief Product Officer, xiaobai, to elaborate on our AI agent teams.

Thank you, Xinhong. I will share our thoughts, mainly on how we plan to build the AI agent teams going forward. Firstly, for e-commerce-related content creation, we've been thinking for a long time. We think that we're not limited to e-commerce providers providing integrated method, but rather helping the suppliers to deliver more tangible results.

According to what we observed, many of the e-commerce agents can design well-designed photos and images, but that's it. Many e-commerce companies lack the sufficient judgments of market and business, eventually causing the resulting product sales are often falling short of expectations. For instance, on Amazon, a product such as water bottles experience intense competition. For experienced e-commerce providers, we need to avoid the oversaturated competition of product background, such as home and living.

Instead, they would pivot to fitness, to outdoor and other high-growth potential segments. So based on this know-how, DesignKit is building an e-commerce design specific AI agent teams. And this new agent team, aside from the original experts that have a good command in design, we have additional roles, such as for the verticals of e-commerce, we have market and data analysts. We have user research experts, copywriters, prompt designers, et cetera, delivering high-quality visual and also establishing a more refined social media commerce video expert team.

For instance, the AI marketing strategist, they can serve as Creative Director, helping users to know what to capture, what to market, or we can have AI commercial directors responsible for the storyboard coordination, helping users to address how to present the visuals. We may also have AI audiovisual model experts acting as photographers, lighting specialists, actors, et cetera, helping users with their contents and their materials. Eventually, we have the AI smart editor compiling all of the editing and the distribution and the post-production methods, helping the users to package and publish their products.

So what we aim to do for the e-commerce is to transform the visually appealing images to images that drive sales and to e-commerce videos that boost conversion rates. The DesignKit's new agent teams aim to help e-commerce service providers drive and achieve real business results.

In addition, aside from DesignKit for Kaipai, for the talking videos, we also hope to create high conversion rate video contents. Based on this goal, we divided the three AI agent teams to assist with our eventual goal. For instance, we have the AI operations team. It helps with the trending topic research, account diagnostics and also content strategies. We also have the AI filming team, helping with the teleprompter assisted filming, the AI avatars, et cetera. We also have AI team helping with the rough and eventual cuts and editing.

The entire workflow is coordinated by the AI director agents. Users only need to submit their industry know-how and their demands, and then the AI agent teams will take care of the rest of their work. For example, when an insurance broker asks for a short AI-generated video about pensions insurance, then once receiving the submission, the AI director agent will automatically break down the information and assign it to the corresponding teams, as mentioned before. And within a short time period, it can provide a ready-to-publish marketing video, and it can even generate social media covers and prompts alongside.

So for small and medium-sized merchants, the advantage of this AI operations team is a 24/7 content providing team. And we aim for niche sectors, for instance, insurance, beauty solutions, local services, sectors that are heavily reliant on social content for customer acquisition. This not only lowers production costs, but also provides a reliable AI team that can have a high conversion rate, so we can help the suppliers to focus more on creative judgment and business operations.

The production handed over more through directions that we have been doing. We have some progress over the past couple of months. For instance, within DesignKit, we already deployed our 4 agents, the e-commerce trial, poster and video agents. From December 2025, the agent function has become a primary driver of our revenue growth on both web and app portals. Our agent has a stable output quality. For instance, DesignKit, web-based DesignKit, it's e-commerce agent roughly has a 50% save rate, where users save at least 1 output per 2 outputs. Now furthermore, once the agent capabilities are integrated into Kaipai more, user experience continues to improve.

According to the latest figures, the penetration rate of Kaipai agents have increased to approximately 11%. At the same time, the RoboNeo, from its official launch to current, it has topped the app stores across 26 countries and regions, demonstrating our globalization potential.

Last but not least, as you may have noticed, we recently launched the Meitu CLI capabilities on our AI open platforms. Meanwhile, the first group of Meitu AI skills have been officially integrated into the OpenClaw ecosystem, covering personal and commercial use cases in the global photo and video industry. We focus on the core skills. We hope we could have from mature technologies around 8 core capabilities and we transform them into reusable skills.

Based on the CloudHub and ACI, we can use Meitu AI's capabilities for photo and video production. We believe as OpenClaw community continues to grow, in the future, users do not need to switch between products. They simply need to have a single command and complete end-to-end creation. This can substantially increase productivity.

Now I hand over to Gary to talk about the 2025 financial condition.

Thank you, Xinhong and Xiao Bai. In 2025, centered around the two core strategies of productivity and globalization, our operational performance continued to improve with sustained user growth. As of December 2025, our monthly active users grew 3.8% year-on-year to 276 million. Paying subscribers grew 34.1% year-over-year to 16.91 million. Over the past year, we continue to deepen the productivity strategy.

As of December last year, paying subscribers of productivity tools grew 67.4% year-on-year, primarily driven by DesignKit and Kaipai, among them. Paying subscribers from international markets doubled, benefiting from the rapid growth of Vmake. Subscription rate for productivity tools reached 9%, an increase of 3.1 percentage points from December 2024. This further validates that with the continual improvement of our products, the output and the output quality, users' willingness to pay is also increasing. At the same time, the increase in subscription rate for productivity tools will help drive the increased proportion of high ARPU paying subscribers, providing support for our future revenue growth.

On the globalization strategy front, 2025 was a year of significant milestones. Our MAU from international markets continue to grow, primarily driven by products for leisure. In the second half of 2025, we launched several global viral features, including 3D figurine, AI group photo, and AI snow. Among them, the AI group photo feature helped Meitu app MAU reach a record high and attracted over 3 million new users from the European markets. It also helped Meitu app top the U.S. iOS category chart for the first time. Driven by these features collectively, Meitu app topped the overall app store charts in 52 countries and regions, and the category charts in over 110 countries and regions.

With the rollout globally for products, we observed in the second half of last year, paying subscribers in the international markets experience accelerating growth with the majority of new subscribers coming from high ARPU regions with mature subscription habits such as Europe, Americas and East Asia. Recently, in the top 50 global GenAI mobile apps list published by a16z, we ranked first in the photo, video and design category with 4 apps on the list, further solidifying our position as a globally leading AI photo and video application company. Looking ahead, we will continue to strengthen international markets. At the same time, we'll continue to launch highly effective features and optimize recommendation mechanisms, further enhancing users' stickiness.

As a result, the oversubscription rate increased to 6.1% in 2025. Currently, the improvement in subscription rate remains our core growth drivers. Besides, with the deepening of our two core strategies, ARPU growth will also gradually become an important growth engine in the coming years.

Before I begin sharing the company's financial results, I would like to note that in November 2025, we have officially discontinued the cosmetic supply chain management services, previously under the solutions for beauty industry. Under the IFRS, this business has been classified as a discontinued operation. Therefore, unless otherwise stated, all financial figures and the year-over-year comparisons for FY 2025 presented hereafter are continuing operation basis.

In 2025, overall revenue grew by 28.8% year-on-year to RMB 3.86 billion, first largest revenue. Our core business of photo, video and design products grew by 41.6% year-over-year to RMB 2.95 billion. Looking at performance by market. Revenue from international markets grew by 37.4% year-over-year in 2025, outpacing growth rate in Mainland China. Revenue from international markets accounted for 38% of total in 2025, an increase of 2 percentage points compared with the same period last year.

From the product perspective, in 2025, revenue from products for leisure and productivity tools accounted for 81% and 19% of overall photo, video and design products revenue, respectively. Advertising next, in 2025, revenue from this part was RMB 840 million, largely in line with 2024. We continue to remain cautiously optimistic about this business, although it is worth noticing that we have pioneered AI-powered creative advertising formats in brand advertisements. Collaborating with international brands such as McDonald's and Visa, this creative solutions have effectively enhanced the social interaction between users and advertisers. Last, revenue from others, which are mainly legacy and non-core businesses, was approximately RMB 62.11 million in 2025.

Moving to the cost side, the total cost in 2025 was RMB 1.02 billion, a year-on-year increase of 42%. Specifically, the largest cost item was revenue-sharing fee to payment channels. The cost increases in line with revenue growth from photo, video and design products, rising 43.5% year-on-year, approximately are RMB 620 million. The second largest cost was computing power and cloud-related costs, growing 16.4% year-on-year to RMB 230 million. The third is third-party API costs accounted only middle single digit of total cost of sales because we were able to use the vertical-specific models and fine-tune open source models to fulfill the vast majority of our user demands, reflecting our in-house AI imaging companies.

For gross profit, because we discontinued the low-margin cosmetic supply chain management business in 2025, the financial data in 2024 has been restated accordingly. Gross profit in 2025 was RMB 2.84 billion, a year-on-year increase of 24.6%, corresponding to a gross margin of 73.6%. The restated gross margin for 2024 was 76%. On a like-for-like basis, the gross margin decreased slightly, primarily due to two factors: one, the change in the revenue mix. Because the advertisement with high gross margins have decreased slightly and because to support our core imaging businesses and design businesses have continued to grow, the computing and API-related costs have increased compared to 2024. However, as mentioned earlier, thanks to our model-agnostic strategy and proprietary vertical model capabilities, the gross margin of our core business remains healthy and stable and will remain so in the future.

On the expenses, selling and marketing roughly accounted for RMB 600 million, a year-on-year increase of 25.5%. The increase in expense was primarily directed to the growth of productivity tools in China and products for leisure and international markets. If we include the portion of combined revenue from photo, video and design products and advertising business, this ratio remained stable at 16% in 2024 and last year. This is consistent with the statistics -- the range, we previously communicated to the market in our earnings.

The R&D expenses were RMB 950 million, a year-on-year increase of 3.8%. The relatively modest increase was primarily because the foundational model training costs decreased in 2025 for two reasons. Firstly, our MiracleVision model foundational training was completed roughly in 2024. Secondly, as we're fully committed to the model-agnostic strategy with fine-tuned open source models, proprietary vertical models, and third-party APIs, we delivered high-quality output for users and did not need to invest significantly into training foundational models.

If we exclude the expenses related to foundational model training, the overall R&D expenses will increase 14.5% year-on-year, primarily driven by investment in our R&D talents. Looking ahead, we'll continue to optimize resource allocation between proprietary and external models and increase investment in vertical model training and R&D talents, continuing to enhance our product capabilities and competitiveness and delivering the users' diverse needs with precision. Administrative expenses were RMB 450 million, a year-on-year increase of 14%.

On profitability, driven by sustained revenue growth, gross profit grew faster than operating expenses. This increased our operating leverage, further enhancing overall profitability. For ease of comparison with last year, the adjusted net profit includes discontinued operations. The 2025 adjusted net profit attributable to owners was around RMB 970 million, year-over-year increase of 64.7%. Under IFRS, our net profit attributable to owners of the company was around RMB 700 million in 2025, lower than around RMB 800 million in 2024.

This change was primarily due to two factors. First, in 2024, we completed the disposal of all cryptocurrencies, generating a onetime gain of RMB 640 million, which elevated the prior year comparison base. Second, as we completed the issuance of convertible bonds to Alibaba at the end of 2025, we recognized a onetime noncash expense of approximately RMB 510 million under IFRS. This expense did not result in any actual cash outflow. Both items are non-operating in nature and are not directly related to our business performance.

Lastly, I would like to share with you the latest progress on the Alibaba strategic investment and cooperation. On the business side, we are advancing collaboration across multiple fronts, including large model, e-commerce and cloud computing supplier offerings. In particularly on large models, we have deeply embedded, the capabilities of Alibaba's open source model series across multiple scenarios such as image editing and video generation to deliver better user experience. Meanwhile, our technology team and Alibaba's team have maintained close communication on the model training. Finally, as the AI industry continues to evolve rapidly, starting with this year, in addition to our two regular earnings releases each year, we will provide quarterly updates on key operating metrics for our core business. We believe this will further enhance transparency and help the market better understand our operating progress.

In addition, we just announced a share buyback plan of up to HKD 300 million, valid for 1 year. This reflects the company's confidence in its future development, as mentioned by Xinhong. The market would consider large model might replace all AI applications. It's obvious in our industry that this might be a misunderstanding. Our growth is also very rapid. So I believe at this time point, it's proper and good for us to do the buyback, plus the 40% proportion. The overall cash return is around 60%.

Thank you. Thank you all for participating today.

[Operator Instructions]

I'm from [ Pana Internet ]. Congratulations on our performance. Based on the latest statistics, the paying subscribers are slightly lower than the first half of the year and the MAU is slightly lower than those in June. So if you could give us a brief reflection about the AI fragmentation, what are its impacts on your businesses? That's the first question.

The second question is how can we better tackle with this decrease in terms of the statistics?

Actually, we have another perspective. If we put statistics aside, the increase of model capabilities has the influence as positive. It can help us give out some products that develop faster. In terms of DesignKit's growth in the second half of the year, there are positive outcomes. For instance, the AI agent functions. Once it's been rolled out, it gradually spread to all businesses of DesignKits ad was taken up immediately. For Kaipai, for example, subscription penetration rate is really high. And the ARR [indiscernible] abroad is also increasing rapidly.

So for us, we actually think from a longer perspective, from a business perspective, the positive outcome of AI agents is immense. As for you, what you saw in terms of the MAUs slight decrease, we've seen similar trends in past years because we need to bear in mind that the MAU is a snapshot of the past month, statistics of the past month. Last year, we had many blockbuster functions that came out in different months. So in December, we happened to have no blockbuster functions. So that's why there is a discrepancy between the statistics and our overall performance.

For the actual numbers, the decrease in growth rates, to me, it's not caused by AI or large models or agents. But rather, if we look at it from a Chinese perspective, the products are maturing. So the growth rate with a large proportion and the large foundation would be slightly slower than before. But this is a positive signal because the productivity products or tools worldwide are growing fast. And these kinds of products are at the early stages of their product cycle.

Put it in other words, in the second half of 2025, it seems our growth rate is declining. But we need to bear in mind that the products development are maturing from the early stages to the later stages. And we also need to bear in mind that the ARPU ratio with the subscription rates are better than we expected. Recently, we tested some new tools. The ARPU of new tools are already at $50 per month, translating to great average income.

So from this perspective, you may think that the growing capabilities of AI is leading to issues. But from our side, we think that the positives outweigh the negative signals. I'm open to suggestions.

We'll move on to the second question.

I'm from Morgan Stanley. Two questions. First, I would say about the expectations for the year 2026. Second, agent is integrated into our products. How do you refer to the subscription contribution and what do we have metrics to measure its contribution?

I'm not talking about too detailed numbers. Also, what I've said before, the overall rate growing in 2026, the top line is basically the same with the 2025, but the underlying or its composition will include more productivity tools globally. In the coming several years, we'll also see a faster growth acceleration. This is the second curve being mature and the rising of the third curve. This might be a direction.

The new adjusted net profits attributable to owners will be different. We'll control our expenses better because most of our operating expenses are about staff and personnel. We are actively tackling this problem by using AI to enhance our output. Meanwhile, with the rapid growth of operations, we will not increase the expenses for staff. I think these are positive directions. More about the agents. The final outcome, we felt happy for it because it solved the long-tail needs of many consumers. For example, the e-commerce, previously, the advertising photo or picture, for example, the barbecue oven, you can only use the outdoor backgrounds in the past, but we actually not only use it, we will also need some meat, smokes, flakes, but it's different. It's hard to achieve in the past. With help of the agents, they can have the chance to express more details. Users can express in verbal language and the AI model can translate it into visuals, representations.

For example, the cat. Especially, in the past, we have cat models, they are very hard to cooperate. We can only change the background in the past. I think we should have a very cute cat on the cat equipment. So with the AI agent, we can have a very cute cat of specific species.

You would also ask, why would you choose Meitu instead of other models? Whether we will be revolutionized by larger models. For the understanding of many vertical scenarios, I think we do better. For example, the cat tower, some cats are not satisfied with what you've chosen for it. So the sellers can better describe the characteristics of their products.

For some nuanced parts, we can cooperate with the traditional conventional agent models or style. The user utilization rate is much higher than the general large models. We can tackle some pain points. Another example, Adobe. Why we think it will be revolutionized by AI, because we know the learning threshold is very high. We need to take -- spend months or years to learn to use it. With the help of AI, it could be very fast. This is also in line with our mission. We want you to feel better using targeted models or the subscription rates or selling.

I think this is the biggest meaning integrating agent and also the point which satisfied our team. For paid subscription, agent is the top line function of all paid functions.

Thank you for your introduction. I'm from [indiscernible] Capital. I would like to ask some questions related to AI. Firstly, as mentioned earlier, the agent function, it has become the main driving factor of Meitu's businesses. So I would like to ask, the agents businesses in terms of the ARPU enhancements, can you elaborate more on the relationship?

The second is, in the future, do you hope for the different vertical sectors to set differentiated pricing for the AI agent or the tiered pricing? Another question is that nearly 19% or 20% of the photo and imaging center cost is taken up by the production tools. So as AI becomes more prevalent in the near future, 3 to 5 years, what do you think will be the proportion of the revenue or production tools in terms of the overall revenue?

Thirdly, what are your outlooks on the gross profits in the future?

Because the ARPU rate is mainly provided in the latter half of next year, there will be fluctuations. But if you look at the DesignKit, we already have RMB combo. We did this combo is because we see there is tangible demand among the consumers. We have other combos at 30% to 35% in terms of the pricing. So the total consumption brought by agents is manifested more profoundly across many products. Of course, the increases in ARPU will need time.

So in the future, you may see that different products will have more pricier subscription choices. Additionally, in the future, there will be different tiers of basic substitutions. But in different sectors, it will be based on the token consumption. In our prices, we discovered that there are drastic different token consumptions. So we use the appropriate skills choice. For instance, if we do consulting, we know that the consulting prices of different sectors are different.

For instance, for high barrier sectors, we need to pay more for consultations. And for the meta prompts for instance, we just mentioned, it's based on sector experience. For the senior practitioners, his or her understanding of the sector have more insights, more personalized insights helping companies to stand out among the homogeneous competition. So that's why we do have the meta prompts and all of the field trial and paid consultations, eventually helping us to gain more insights into how to develop our vertical sectors. So eventually, we base on the different sectors different values and providing more returns for our consumers.

In terms of the gross profits, we dropped by 1 to 2 percentage points, but the greatest reason is because of the changes in the revenue mix. Because in the past, advertisement took up a lot. But as advertisements proportion continue to decline, in the short term, the gross profit will decline as well. This year, the advertisement proportion will decrease, but it's already taking up a lower proportion. So its influence on the gross profits will be lower compared to 2025. Additionally, this year, because Apple in China is reducing its fees in China, it's helping with our gross profit.

As for your concern, I think you may be concerned about third-party API use and deployments. You may think it influence our gross profits. But at present at least, because we only have a mid-single digit of our third-party API so it won't influence our gross profit that much. Overall, our gross profit will remain stable this year, around higher than 70%. We haven't had the specific statistics for the proportion of production tools in our overall revenue. If we look 5 years into the future, production tools will take up a higher revenue proportion than leisure tools.

Put it in this way, production tools as discussed for 2.5 years or 3 years from 2023 until present. Now it takes up around 19% of our overall revenue. But actually, for the vertical sectors, we managed to integrate industry know-how into the products through the AI agents. So there is a scenario that I believe in the future, I think the growth rate will be different from 3 years before. It will be faster. But as for the specific statistics, I have yet to collect them.

Thank you for your question. We will continue with the Q&A.

An analyst, Sara. I have two questions. First, as mentioned by the management, the cooperation with Alibaba especially for the products for leisure, what about the user portals who will have more AI-powered phones that would be integrated with many AI-powered functions. So how we see the changes in user portals for products for later?

Question two, the productivity tools ARPU is more valuable or promising than of the products for leisure. We are the model-agnostic strategy. We want to build the model to tackle the pain points of users. The models are evolving rapidly. How does the management view the extension of models, this possibility into the photo, video and design vectors? For example, will the foundational large models be extended into the e-commerce area or sector?

Products for leisure, from the perspective of user, I think we have an advantage because we have an MAU of over 280 million. Many new products of ours published in recent years have performed very good with rapid growth for DesignKit and Kaipai and also Vmake, Wink, we have access advantage. So we can be top-tier players in this sector.

Other giants, including the smartphone producers, they, of course, will try new functions related to photo, video and design. It's always a very essential part of them. For example, many smartphone producers focus on the imaging capabilities of the phone.

Of course, we feel the challenge. But we will find many corresponding room where they left or they overlooked because there is no one company that could serve the users' all needs. This is a marathon for innovation. We view this trend very objectively. This challenge, in turn, makes us feel the warning and have the spirits to challenge is good.

And second question, what about the extension of large models into other sectors, I think the answer is yes, but it's similar to the answer of the first question. No large model companies can solve all factors on use, especially vertical sectors pertaining certain areas. A large amount of energy and expense should be devoted. So at present, we are digging into these perspectives and we saw the things that these models could not do. I believe this is also another marathon of innovation, different ecosystems for models and applications.

Models cannot cover all applications. Actually, it's empowering applications. We know we would do something about it. They develop some independent targeted apps or software. For our team, they make us feel pressured and also urge us to increase our competency.

I'm from Citibank, Internet analyst, Vicky. I have an AI-related question. I look forward to your thoughts. I don't know what are your thoughts about this.

Indeed, we saw this trend as well. Xiao Bai has elaborated over the AI teams. How we build the teams and how to deliver high-quality results, we also mentioned earlier that the business models have been changing from the subscription to the consumption of tokens. So to deliver high-quality results, so we think there's an opportunity for us. I suppose we roll out some new production tools that have a relatively high barrier, but users are willing to buy because they can deliver high-quality results.

As a result, we see our ARPU rates increasing fast, especially production-related. Therefore, in June this year, we will post the Meitu Imaging Festival, and that will be launched, focusing on the AI agent teams' establishment and how to deliver high-quality results for our users. I believe, by then, we will have a better understanding.

I would like to add some more thoughts. I think an interesting thing is that there are many store apps through the subscription revenue. So after you subscribe to something, you may not use it for 99% of the time. So that is an issue of overpayment. It can't be like this. So there are many inefficiencies in the previous conventional model. But this cannot be applied to us because, in the past, we've been providing services that needs to be subscribed to be used. We don't have a model where we subscribed and the users do not use it for most of the time.

So even though some people regard us as traditional imaging and photo industry, but we actually have some differences in terms of the conversion rate. For the subscription, I think the token consumption, a part of our revenue will be generated by subscription. It's just that in the past, by subscribing, we provide a group of tools. But now by subscribing, we have a bunch of tokens. After you consume all the tokens, you may need to buy higher-tier subscriptions that gives you more tokens. I think for the two models, they are not dichotomous. They are a combination.

Due to the time limit, we will continue with the remaining two questions. Before online questions, we will have another question on site.

I'm an analyst, Xixi from [indiscernible]. My first question is that globally, last year, when we're expanding production tools, we mainly focus on vertical sectors,; for instance, the restaurants, et cetera. I wonder, in 2026, do we expand our vertical sectors to more sectors?

Secondly, when we look to the production tools of users abroad, their habits may be different from those in China.

Thirdly, I would like to ask more about your cooperation with Alibaba. Alibaba's users in terms of production tools, what are the percentage in terms of the contribution to the revenue? In 2026, the driving of production tools mainly come from ARPU or the paying tools.

We're now doing extensions in vertical applications. As you may have seen, you have saw the financial results video and some AI agent studio functions. Besides what's mentioned, we saw very vertical sectors where there are opportunities for us. Under these scenarios, traditional services, they may spend over USD 10,000 to USD 20,000, but we can achieve this with 10% to 20% of that with same or even better quality. There are scenarios like this.

Meitu will try to avoid the directions of manage giants, their core paths and strategies, so that we would not be greatly influenced. But we'll try to discover the parts where they may neglect or didn't consider it very important. So actually, we have many opportunities to lead in very vertical sectors. So we may try to avoid some popular or hot choices. For example, we still have a large user base in the off-line market. Traditional expense is high, but now it's just maybe 1% of the previous expense. In detail, I can't expose too much, disclose too much. But you will see our future products.

And about the user habits of the productivity tools in international markets, we may observe different vertical scenario preferences and. Last month, our RoboNeo topped in the photo video category in 15 countries, including Brazil. For example, the South American users, they have a strong desire to express on social media. This is different from Asian markets.

And we'll do more language and operation support in Southern American market. We cannot express too much about it, but I want to say that we have different orientations and directions. We will segment very carefully and precisely. We will try to serve well. a certain need and extend to more categories and markets.

As an additional reminder, as what Xinhong have said, the critical point is that the users demand on visuals because there are different cultures, they're vastly different than us in the past. But now due to the cultural differences, we need to have many different and AI adaptation. For instance, the same e-commerce platform, Chinese users may perform tests. They have less coherence but have more visual stimulation. But for overseas users, they may prefer more simplified accounts that look more high end. So there are great differences.

For instance, Brazil, many Brazilian users would like the e-commerce platforms to stand out more. Some countries like light colors, for example. In our main business areas, we have all the compatibilities and adaptations. After the adaptation, we also observe the different subscription habits. For Chinese users, because we would like to subscribe low-cost orders, in China, the per week subscription is more preferred than monthly subscription. But for the U.S. market, the reverse is true because they think weekly subscriptions, it's is hard to experience the value of products in the short run. We need to have a longer time period to understand what values can the product bring. Because they are less sensitive on the price, they hope that they would pay to create value and then decide whether to subscribe.

So for us, in the past, many of our products were based on the demand and habits of the Chinese market and then we ready it to the rest of the world. But now we hope we cater to the specific demands of each area in the world. In the past, we may have 100 products in China, but only 50 to 60 overseas. But now we look at the overall demand, and we have products that are 90% available and also 90% available in China and also overseas. That is our biggest change.

And due to the time limit, we will have the last question. If we have investors on site that have questions, please raise your hand. So we'll have a question online from the meeting.

First, about the ARPU value. We mentioned we can see increase of the ARPU value. The main drivers are productivity tools in Chinese Mainland and international markets. And also, the third-party CPI are single digits. Which directions we use the third-party API? And why would we consider this of high, low status of proportion?

Management answers the elevation of ARPU value, the main driver is that the introduction of agent. Many delivery of outputs will consume tokens. After the consumption increases the purchase more expensive packages, due to high-quality, they do not mind that much because this cost is much lower than in the past, nearly 10% or 1%. So the main ARPU growth are contributed by this partner, productivity tools.

Products for leisure, I will not consider a dramatic increase in ARPU value. But with the higher proportion of international markets, this part certainly increases, but still not that huge.

About the API part, compared with other model companies, you know we pay attention to quality or effector. We have an evaluation team of over 200 people. Many designs are outsourced. They comply to teams, but ultimately it's up to R&D team. But in Meitu, they are competing. If the design team do not consider it good, it's not published or aligned. For example, Xinhong himself also paid great attention to it, one, because the evaluation of the effect of different models.

When we come up with a case, we first evaluate all models about their effects and outcomes. For example, 10 models, the winner, does it meet the expectations of the real users? If it's good, we use then the API. But after the competition, if we cannot find a very good winner, we self-develop. So our understanding about the external API, first, we evaluate. And if it meets our standard, we use it or we choose to self-develop. We hope they cooperate and build the industry's top standard.

Thank you for your questions of investors and the detailed response. If you have more demand for our tools for statistics, you may contact the IR teams. Now we'll enter the media Q&A. This session will not have simultaneous interpretation. For the participants on Zoom, you may feel free to log off.

Thank you for your participation and attention once again.

[Statements in English on this transcript were spoken by an interpreter present on the live call.]

Good afternoon, everyone. Welcome to Meitu's 2025 Investor Day. I'm your host today, Coco. Today's event is divided into 2 parts. Firstly, we will demonstrate several [Technical Difficulty]. Secondly, we have invited Mr. Wu Xinhong, Chairman and Group CEO; Mr. [ Ngan King Leung ], CFO; and Mr. Chen Jianyi, CPO, to engage in the discussion with everyone.

Without further ado, let's begin the product demonstration. As you know, we categorized our products into products for leisure. and productivity tools based on users' needs. Today, we'll showcase the products for leisure, Meitu App for photo editing, Wink for video editing. And under productivity tools, we'll present X-Design for e-commerce material design and Kaipai for talking video creation. Finally, we'll demonstrate our AI agent RoboNeo.

First, let me demonstrate the Meitu App. Your impression about it tends to focus on portrait beautification, but it delivers excellent results in image editing scenarios. Let's have a basic look, an example for you guys. We just passed our National [Technical Difficulty] you felt the crowd. Suppose we are here at a tourism attraction [Technical Difficulty]. We have many elements and factors. So we cannot have a photo focusing on the main character. Now we want to remove all those elements in the background. Now we can click the button in the middle, the removal brush. Based on traditional function, we can use classical removal function to just eliminate some elements in the background. For example, the people, but you will find some bigger objects, may not be that clear after removal because we do it based on the stretching pixels. You will find the blurring signs.

Now let's look at the right part of the picture. We use the GenAI-powered removal function. We removed some elements we don't want. This technology, first, it recognized the objects we want to remove. And then based on its understanding about the background, it generates something similar to original picture. [Technical Difficulty] free function, but the [ write ] function needs to be paid. So this example can tell [Technical Difficulty] subscription instead of free options available because the effects are better. This is also one of the contributors for us ranking in the top 5 of many app stores.

Next the picture editing. We can see inadequate lighting, we can see the dark parts about the face and body of the model. We can improve the lighting effects. First, the backlight. One of the main issue is that the overall brightness is dark. Traditionally, we just adjust the overall brightness, for example, like this. But you will find after we improve the overall brightness, we do not have the quality or the atmosphere of the magazine photos because some shadows that should be dark are also brighter, very satisfactory.

[Technical Difficulty] GenAI, the AI backlight after generating. It just imitates and simulates our natural lighting. For example, some areas should be brighter, maintaining darkness for the dark area -- should be dark. You can see the final result. After comparison, you can easily find the function, powered by GenAI, can create better quality.

Besides what we mentioned above, powered with GenAI, [Technical Difficulty] your younger self or with a friend from far away or your childhood with your parents' childhood photo together. Actually, I can do this. Recently, we launched a new function called AI Group photo. You can combine 2 models very naturally [Technical Difficulty]. Based on this case, I select these 2 models. Due to the time limit, we generated many samples for you guys to refer, like this one. This creative function powered by AI quickly drives the globalization of [Technical Difficulty]. Lately, it is very good [Technical Difficulty], top in 14 European countries' app stores and also in many other regions. We will further propose more creative functions to attract more clients from different regions and countries. [Technical Difficulty] Above is the photo editing for the Meitu App.

Now let's turn to a function of portrait beautification. Before we share some paths of our globalization, we would like to find some common pursuits of global consumers. Let's look at this picture. Actually, globally, our clients want to do something related to anti-aging. We believe this is the common pursuit for global users, despite your regions, we can see very clear wrinkles. [Technical Difficulty] editing leads to improve the aging situation of this model. We can also use some free functions of many steps to do so. But today, we try some very simple [Technical Difficulty]. We just do it with one-click facial plump, adjusting the lighting on the face to lighten the wrinkles on the face.

We also use the Even and Lift function for paid subscribers. We adjust some muscle directions [Technical Difficulty] or around the mouth and lips compared with the original one. For example, we can use this high-definition [Technical Difficulty] multidimensionally compared with the original picture. We use 3 steps. We can replicate the natural effect. Let's look at a sliming example. This picture actually, it doesn't have many flaws. We want to have better details around the model's waist or hips. We can do some minor adjustments. The background details are abundant in this picture. If you want to adjust [Technical Difficulty] powered by GenAI, we achieved background protection. Before we adjust the waist, we protect the background first [Technical Difficulty] but you still find these users from the Euro-American region tend to be more about fitness, so we can have clear app lines. It fits the aesthetics of the Euro-American users.

We can turn to some niche functions. But once you have this kind of demand, you will pay for it. This girl with braces, she wants to smile brightly. For myself, I also did braces when I was young, but I refused to smile very brightly. Powered with GenAI, we have teeth adjustment function. We can remove the braces just with one click. And we can also generate teeth adjustment -- alignment [Technical Difficulty] demonstrated [Technical Difficulty].

Coming up, we will focus on video editing and using different differentiated marketing policies and our own policies, our Wink product. It has been a fast-growing product. It has 2 core functions. One is quality restoration. The other is video portrait beautification. Today, we'll demonstrate these 2 core functions. First, we open quality restoration. Firstly, we restore a photo. Suppose we choose a photo here. To the left was the before demonstration. You can see that the entire photo is quite ambiguous, low resolution, and the water coming out of the holes is unclear as well. So this is the quality restored photo. We can see that almost every detail of the photo is now clear. And the lighting and details around the model's face has become clearer as well. And if we use another example of a child, we could see that the original photo, I believe many investors here, when they took your photos as a baby or as a young child, the limitation [Technical Difficulty]. So many of the old images that we may have tend to be clear [Technical Difficulty] by Wink.

Actually, the AI here, basis its restoration, on the analysis of photos [Technical Difficulty] mainly focused on video editing. But because we have video portrait beautification, aside from videos, photos, restoration is actually a big hit function. So that's why we have both developing. So that was quality restoration for photos.

And now we'll see the video [Technical Difficulty]. This is a cartoon from 1961. This was the original video. So 5 minutes needs 54 credits in our Wink and [Technical Difficulty] more credits. So now I select this video. And because of the time limitation, we have a sample of this restored video. You can see the before and now after [Technical Difficulty] of restoring video of a live streaming vlogger. This was the before [Technical Difficulty].

And now we move on to the after, the GenAI-powered restoration technologies based on improvements, the [Technical Difficulty], the overall video development. So now we look at an example of GenAI-powered video restoration in a dimly lit environment. [Technical Difficulty] is situated [Technical Difficulty] -- so we see the before effect. And now we see the AI-powered [Technical Difficulty]. So you can see the effects are quite stark.

Now another key core function of Wink is video portrait beautification. [Technical Difficulty] no makeup on. Now we begin the video portrait beautification. Firstly, I would like to [Technical Difficulty]. Suppose we choose [Technical Difficulty] on the vlogger's face with some Halloween example, so we can have some plumping for the faces, and we click on the plumping function. And we also add some lighting and widening to the vloggers' facial details. We adjust the parameters, and now we de-wrinkle -- we reduce the wrinkles. And then we add more details to the facial features of the vlogger. Because the video is quite short, we can add more details, such as makeup. We could -- if we remember the effect without makeup, now we see the effect with makeup.

Above are the core demonstrations of Wink and the Meitu App. So now I'll demonstrate to you the newly rolled out AI-powered [indiscernible]. So now before demonstrating, I would like us to think about our past experiences of shopping online. Did you encounter any unpleasant situations? For instance, have you received clothes purchased from online and became dissatisfied? Why? Could it be because that your physics didn't exactly fit into the design of the clothes, or the textures or materials did not match the product image [Technical Difficulty], or was it because you lack extensive fashion styling knowledge, so you simply couldn't replicate the stylish outfit combinations, or you have a special or important occasion and you do not know how to choose your outfits? Or sometimes simply your wardrobe is too full. You don't know where to begin with choosing your outfits.

So now based on the above pain points of our clients and [Technical Difficulty] solution for online shopping. So now suppose we opened the app and logged in. Firstly, the app requires us to put in some basic information, your gender, your clothes, your wear, your height, et cetera. And your size, you may be able to wear regular size. And for the size, we provide some regular body shapes -- physics, the hourglass shaped, apple shaped [Technical Difficulty] are the hourglass shaped [Technical Difficulty]. Now suppose we choose sportish traveling, the vacation -- on vacation style. And for the regions, perhaps we choose a K fashion. And then for the artistic styles, we choose minimalist dressing style. And we see this is the AI-powered recommendations. These are what the AI recommend us to purchase.

And if we select the first set, this is a yellow and cute outfit. So we could click and see the specific details of each of the outfits. Suppose we click on the yellow vest. Here, we see all the details of the vest, for instance, the sleeves, the shape of the color, the patterns, the color and the seasons, et cetera, all of these fundamental details. And in the recommendation, there would be Taobao links. You can click on these links, and you could buy them in no time.

Now we return to the product itself. So aside from recommending based on consumer preferences on the homepage, there would be [Technical Difficulty]. Suppose the user chooses a recommended styles, they could save it in the outfit shopping list, for example. For instance, we could save it for tomorrow, and the system will remind you what tomorrow's outfit will be. So aside from these recommendations, the user can actually upload photos of their existing wardrobe -- of their existing outfits. For instance, [Technical Difficulty] now we upload the photos. The app will scan and distinguish the specific garments that we have. It takes some time.

So we can see that the generated result is the exact replica of the clothes in reality. The length of the sleeve, the texture of the color, et cetera, are all exact replicas. And we can click the add to closet below, and the system will distinguish and sort the items for you. For instance, the clothes that we just identified belong to topware. And this will be -- this will make it easier for clients to sort their clothes later.

And the system also generates descriptions based on the uploaded photographs. And the AI system can categorize clothes based on their conditions such as they are wearable, they should be disposed, or they should be put on sale, et cetera. So now we see that in the bottom of middle, there is a function called [ WOW AI ]. In fact, there will be a shortcut of this at the homepage. This allows users to communicate directly with the AI agent to do AI trials, which is what we will demonstrate just now.

Now suppose the user has a significant job interview, and he or she needs to buy outfit some clothes [Technical Difficulty] that I should wear for a job interview. And now the AI assistant can think based on the users' scenarios and generate outfits matching combinations based on the users' preferences. So now we see the recommended combinations. If a user on the [Technical Difficulty] and now the user can upload a full-body or a half-body portrait of the user. And now the results -- the users [Technical Difficulty] user satisfied, he or she can click on the bottom-right link and see more details of specific outfits. And then the links will also lead the user to Taobao, and they can purchase directly. For instance, these are the shoes that have just been recommended.

And now another scenario is that [Technical Difficulty] going into winter, and the users need some beautiful yet warm clothes. So now we put in "What should I wear when the weather becomes cold?" So now AI, again, will generate recommendations based on the users' preferences and the weather conditions provided by the user.

Now we have -- based on your uploaded portfolio, it will generate the try-on effect generally. You don't need to regenerate. If the user is satisfied with it, you can just click on the picture to go to the top item link to it. Above is the introduction of where WOW, in the future, we will improve the whole product, the try-on effect and the user experiences. We want to achieve stable effects for generation and also faster speed or for more clothing types.

We just demonstrated for clothes. We want to do jewelry, hats, scarfs and all of those accessories. We want to achieve higher consistency to replicate the true cutting of the clothes and how to fit clothes on different body shapes. To do more detailed and realistic representation [Technical Difficulty] or how to match clothes together.

Next, let's welcome Irene to demonstrate productivity tools, X-Design, DesignKit and Kaipai.

Thanks, Coco. [Technical Difficulty] X-Design, it's for the e-commerce to do SKU designs. You use it on the web or PC. These are the common scenarios. I believe you have a basic understanding about X-Design. So in limited time today, I want to demonstrate some new functions.

You also see besides the traditional homepage, we have an AI agent program [Technical Difficulty] integrate this function into [Technical Difficulty] might ask how AI agent can empower e-commerce practitioners? First, let's look at the traditional core functions of X-Design. They are based on [Technical Difficulty] they just do some upload. But in the design process, it didn't engage a lot.

Now we have a lot of user cases. The agent can help the e-commerce practitioners to design. Now the e-commerce, they can be a designer for himself. They don't need to pay for outsourcing designers. [Technical Difficulty] show relevant functions. For example, do the refined [Technical Difficulty] some e-commerces, how many pictures do I have for different platforms like Amazon and Tmall? What angles? What pictures? Now with the help of AI [Technical Difficulty] major use cases based on our investigational users' needs, it unleashes users' product [Technical Difficulty] help to generate a series of relevant photos.

So now let's just demonstrate it. Suppose today, we sell womenswear. We have a very basic SKU. It's the [Technical Difficulty] with AI about it. For example, I want to do SKU variances. I want to do 3 colors, light pink, dark green and leather black. After the conversation, RoboNeo starts to imagine. So we had a complete conversation sample.

First, RoboNeo analyze loaded [Technical Difficulty]. It says a jeans jacket and also the details, whether the background is transparent or whether the background is complex. Based on our needs, it generates 3 colors we want. The light pink, dark green -- this is also my favorite. I can show you the details. The third is the leather black. RoboNeo, it replicates the texture of leather.

After that, if I am just an entry-level e-commerce practitioners, we want to a series of pictures, how many pictures do I have? I just said, "A series of photos for Amazon." First, it will do further and deeper analysis. What scenarios does this picture fit? So RoboNeo considers it is for many street photos and later wearings, a very unique design on the shoulder and also a pocket with zip and some metal [Technical Difficulty]. RoboNeo can all grasp this information.

RoboNeo points some potential points for improvement. For example, we have the pure white background. In the future, it will automatically help us to change the background or to show some trial, in fact, impact [ wave ] models. First, it gives a refinement with the white background and 4 detailed demonstration pictures, including the name for different details and also [Technical Difficulty] including [Technical Difficulty] because it's generated by AI. You can [Technical Difficulty] 3 added the characters. This will save a lot of time for e-commerce and gives you a demonstration picture without model, with light and shadows.

Next picture, try-on effect with model and interact [Technical Difficulty] this jacket is natural. The details are preserved and also the painting in the background and the [Technical Difficulty].

Okay. About the length and the width of this picture and clothes. It will generate a size picture with 2 more details. Next, this is a picture sets for Amazon. It didn't ask for further props. It's very friendly for e-commerce practitioners. But if you want more -- if you want more try-on effect pictures, I want to achieve a picture of try-on effects with rural America model. It will do targeted analysis. photo style, the background characteristics and the poses of the model. After analysis, it will give us suggestions. For example, actually, it gives 2 different suggestions, a same model from the front side and the other from, the left side.

I even want more. I want a video. Then based on this picture, I want it to be live or moving. This is the processing part of RoboNeo. It finds out that this picture can be divided into 2 parts. It gives us more details or respective playing. You can see the 2 different parts are not interfering each other. The model moves naturally. [Technical Difficulty] I want an Asian model this time. The same process for thinking, first, it analyzes and then what's the characteristic of these clothes to ensure the consistency of these clothes. Now we have a picture of try-on with Asian model. The effect is really natural.

[Technical Difficulty] whether it meets our requirement. Sometimes maybe RoboNeo is not sure of your demand, it will ask you to -- for permission to proceed. This is also a way for RoboNeo increase its efficiency. We think it's good. We proceed [Technical Difficulty] scenarios and backgrounds.

First, just like the fifth avenue in France, the ratio is very proper. Second, interior, maybe at home with brighter background. Third, leaning on a cliff in background, in nature is very cool. You can imagine if you really invite a model to do a picture like this, the price can be really high. Fourth, in a runway, classic demonstration. It preserves the integrity of the model and also the details of the audiences, including their hands, their foot, feet. Without further prompts and demands, it could generate demonstration pictures. It asked for expansion or further proceedings based on one but [Technical Difficulty] RoboNeo voluntarily asked for further proceedings.

Now I would want more poses for the railway track. First, the front, second side with hands in the pocket. So you can see a [ batch ] in the pocket. This is really vivid and realistic. Next [Technical Difficulty] whether should we generate a video based on the picture, why not? I need [Technical Difficulty]. It can create a video of 5 seconds based on each pick. [Technical Difficulty] really smooth and natural, basically [ emitting ] on the runway, some basic habits, touching a hair, eyesight to different directions, then it mixes them up for 15 seconds. Based on that user can do some sample editing, and then post it on social media or platforms they would like to do.

Next [Technical Difficulty] do we need a detailed picture? It gives 6 different details. We can see, this is popular in many e-commerce platforms. It knew it is a picture of clothes. So it generates a table of sides. You can also add it in the future by yourself. [Technical Difficulty].

So as the e-commerce merchants, sometimes we need renditions of children's wear. So -- but sometimes finding children to become models cost dearly, and the time can be less flexible because children could be uncomfortable in the shooting frames, et cetera. But now with RoboNeo, we can fix this issue within seconds. So now we ask the AI, please apply our clothes on to a child. So after analysis, RoboNeo asks us what kind of background do we need. So suppose we had a street shot. Now again, we need [Technical Difficulty]. Now we see these are the results of a cute child posing and doing a street shot photo, wearing our clothes. And even the parent with a bag of the children's [Technical Difficulty] more realistic. The AI asks a follow-up asks us if we need another scenario of at home or near the beach. And now we choose at home. So we now have a slightly older child model with our clothes, with a handbag at home. So we can see that while the design [Technical Difficulty] RoboNeo [Technical Difficulty] rely on in-house or third-party image generators.

Now we have an instance of photo, so we will now generate a model holding the incense in her hands or in his hands. So now the RoboNeo asks us what kinds of hands do we need? So -- because my demands are quite simple, I have a simple follow-up instruction. So now these are the generated image of RoboNeo. So whether it is the gesture of the hand or the background, they're all natural and smooth. Perhaps I think the background is too boring. So I would like to have the -- background to be a lavender farm. So now you see the result. And the overall result only took a couple of seconds. But in traditional scenarios, for designers to simply make this image, they could take hours.

So maybe I would like even more. So I would like the image to become a video. And this is within the capabilities of RoboNeo. So we can see the wind blowing over the lavender farm and also the hand slightly shaking while holding the incense bottle, all makes the video more realistic and more vivid. So again, traditionally, this would take hours and cost dearly traditionally. But in our platform, especially in RoboNeo, RoboNeo can do it within seconds with several clicks. [Technical Difficulty] functions of AI agents.

Now we would also have AI generations based on pets, based on accessories, et cetera. For instance, we are a pet accessories producer or merchant. RoboNeo will still contribute greatly because if we use pets to shoot [Technical Difficulty]. So in a realistic e-commerce platform, after designing these set of images, we still need the eventual poster. Of course, we can make the poster one by one, but that would be time costly. So we have a function where we can help e-commerce platforms to produce posters, many posters simultaneously. So we open e-commerce materials. Supposedly, we need a 3x4 e-commerce poster. So we can see to the left, there are templates. Based on different products, we have different templates. So the clothes that we've just demonstrated. This is a kind of an autumn style clothes. So based on the template, we can use some of our own products. And we can see it to the left, there are producing templates in multiple cases.

So now we see [Technical Difficulty], and we add them simultaneously. And we could add some basic information, some basic descriptions. For instance, the prices, the style, the classic denim, the black leather, cute pink and lastly, [ cool girl ] green -- dark green. Supposed we sell it at low prices, CNY 299, CNY 299 for all colors. So the following step we see that all [Technical Difficulty] AI agent to the left.

Suppose we will like the prices here, so we could match the labels. And we have a description here that we would like to ship it to other places to describe what are the clothes. So now we would arrange the demonstrations and all the words. And once we click, we can have 4 pages. It would take some time, but it is still much shorter than the conventional process. So we could see that these are the 4 generated images, and we could still make some adjustments to the results. [Technical Difficulty] So we can see the eventual results. And based on AI generations, the e-commerce merchants only need to make minimal images [Technical Difficulty] that satisfy their demands. So overall process is not [Technical Difficulty].

Aside from this function, there is another function that I would like to demonstrate, which is the AI live streaming marketing video. We could find it at the homepage. So the bright side is that we only need to upload 3 videos. These are the pre-generated videos by AI agents that I [ mentioned ] earlier today. So AI agents can distinguish what kind of elements do you have in these respective videos and make recommendations. For instance, once it sees the denim jackets, it can help us recommend the styles for customers, both in Chinese and in English. Because it takes time, we have some pre-generated live streaming e-commerce videos. We have not edited any of them. These are raw videos and e-commerce merchants can make further adjustments based on these raw videos with the subtitles, for instance, with the BGM, et cetera. So these are -- this is an eventual result.

So some of these defects, we could make further adjustments, and we could do some minute changes. So we continuously emphasize that AI cannot help you make the eventual product 100%. You still need to make the final adjustments. Because if you rely on AI generation solely without any final touches, this is far from enough. So aside from these promotion and marketing videos, the final is the producing product images simultaneously. So these are the 4 images we've just produced. We can see to the left again, there are different templates. And one of the templates are for clothes. So we could produce all of them [Technical Difficulty].

So these are some in-app purchases that consumers need to buy. And of course, we have the eventual results based on the different images, and it gives me 4 choices for me to choose. So as an e-commerce merchant, this is very convenient for me. I can choose any scenario that I want, make some final touches, and I can upload to my platform. So these AI commercial product images have been released in the end of September this year by -- onto the platform of many merchants. And one by one, the subscribers and users of our apps can see these newly produced functions. So these are the demos of X-Design.

And now I would like to -- RoboNeo. And now I would like to move on to Kaipai. Kaipai is mainly based on mobile phones. So I would like to use my mobile phone to do my demo. Over the past 2 years, Kaipai grew rapidly. Based on my experiences, I felt also firsthand that Kaipai combined all the functions that the user needs into a single workflow. So from a user perspective, the overall photography and photo editing functions are comprehensive. So now suppose [Technical Difficulty] with millions of fans, and we would like to demonstrate the -- talking -- the functions of Kaipai.

If we would like to do talking video, traditionally, it would take a long time, at least an hour or 2 to write the scripts and then [Technical Difficulty] would take even longer [Technical Difficulty]. We open it, and we type in our topic today, "Please share with my fans about the product, Kaipai." So now I can choose the word limit, 250 or 500. So today, because of the time limit, I will set it at 150 words. So a unique function is that Kaipai can combine the scripts with the own vlog and uploads of the vloggers. For instance, if the vlogger is a social media vlogger, mainly aimed at sharing good softwares with users. So the scripts would be tailored to my demands. Suppose that I'm this kind of social media vlogger. I save my preferences and my word limit, and I would generate my script because, again, it takes time. So to [Technical Difficulty] the AI, I have pre-generated some -- pre-generated a script. So now we can move on. This is the generated script.

To the bottom right, we can see that the Kaipai is now guiding the user to actually shoot the video. This is my phone doing the demo. So you can see to the top is the teleprompter. So suppose you need more makeup, I could add on makeup using filters. For instance, the changing and makeup to my nose to my eyes, to my lips, et cetera. For instance, I could brighten up [Technical Difficulty] lipstick effects, for instance. There would also be records kept by the AI. For instance, if online celebrities need to have multiple videos in a day, once they open the AI function, they can have predetermined templates of lipstick colors, et cetera, so that they do not need any further adjustments every time.

So after adding the makeup, I think my background is too monotonous. It's only a wooden plank. So whether it is -- so I can change my background to an open window or an office background, et cetera. So now we use this open window background with the tree behind. So once I begin recording, I don't need to memorize the script. I can simply read off the prompter. But you could pay attention that as I read, the prompter is following my voice. It is making adjustments based on my -- the speed of my speech, and it can guide me. So now we begin, "Do you think making difficult videos is hard?"

If we speak a wrong sentence, we can just withdraw from the last sentence. We don't do anything. You can see the proceeding of the subtitle according to your pronunciation. This is the end of the recording. Next, we click the button finish, complete. [Technical Difficulty] We can find the parts which we have spoken wrongly. We just check and delete it. This is the part we just spoke wrong. If it's some unnatural parts, you can redo it. But due to time limit, we go on. We have a very Internet-popular template. It used to take 2 hours to make a video. With this kind of template, it unleashes a lot of productivity for our vloggers. We see this bilingual advanced high-level RAV template. Notice, [Technical Difficulty] tap on the button template. The subtitle would change its style according to the template. The bilingual subtitle are automatically completed.

Next, we can add a very interesting beginning. We know that on TikTok, on RedNote, the first 3 seconds are called golden 3 seconds. It decides whether a user would go into your main page to be your follower. We can achieve some interesting effects. [Technical Difficulty] established some beginning templates. We have a film style beginning. This is a paid function. Not satisfied with one, you can change to another [Technical Difficulty] template. You can have some creative ones [Technical Difficulty] on the bottom of the screen, you can try it by yourself.

Next, how to make a cover for the video? In the past, we select a beautiful frame or even you just screenshotted and maybe you will turn to other applications to do cover [Technical Difficulty] one-stop effect. First one, I can directly change the character, the word, but I personally like the AI one more. AI will based on my previous frame to help me design some popular [ beginning ] cover. You can see the effects really good. For example, I think the third one is good. I choose this. If you think the speed is little slow, [Technical Difficulty]. A simple video is generated within 3 minutes. I can just click the button save. This is why Kaipai grew fast for the past 2 years because it solves many demands for the users.

But what for those vloggers with millions of fans, it records [Technical Difficulty] cyber yourself, you can design your own digital avatar. It will imitate the way you speak, you pronounce, your eyesight and also your movement. It will learn, deep learning, machine learning. But you need to film a very short video for demonstration for it to learn, and you need to authorize due to some consideration of security. This is a premium feature so you need to pay extra for it, not only subscription can do it too. I think it is HKD 250 for it. Due to time limit, I generated my own digital advertising yesterday. One I filmed at home. So it will match -- it will do the lip syncing. You can see the movement of the lip. It follows closely with the actual pronunciation. I click the function, reading the text, we can adjust the tone. I prefer to use my own voice. We can also record a [Technical Difficulty] or recognize our tone. It's my tone, but I didn't speak [Technical Difficulty]. You can see the process I generated yesterday.

You can listen [Technical Difficulty]. You can go to edit, click the button. I also did a complete sample to demonstrate it, [Technical Difficulty]. Including the selection of BDM is included in the template. The overall effect is natural and can save a lot of time for vloggers. They can do more videos every day.

And also a new function launched last month, it is called Intelligent Mixing and Editing. Small shops need to do advertisement videos to attract customers. It may not only be an owner talking, but it will film a lot of the environment, the product, adding subtitles, the product details and so on or some other use cases. For example, in the national holiday, you will take a lot of photos. You can just do a video with just a one click. Yesterday, I filmed several short videos for us to demonstrate today. I'm going to still choose my own voice tone. AI can help you generate the subtitle, the script, but I wrote it yesterday for the Investor Day. Now we can just [Technical Difficulty]. In this process, first, it's still analysis what we have. It needs some time to deal with it. It analyses what is contained in our video, the glass gate, the front desk, the receptionist, the elevator. So you have the movement of walking upstairs.

I want to demonstrate to you the thinking process of Kaipai. In the modern building, we have smart gate, interior designs and furnishment. I already prepared a sample. You can see the overall effect. If you think it's okay, you save. After you download, this is the video you will get. Do you want to use it? Do you want to try it? [Technical Difficulty] What you have in your videos and photos? This is called AI Intelligent Mixing. Above is the introduction about Kaipai.

Next, let's welcome [ Janet ] to introduce RoboNeo.

Thanks, Irene. Next, I will introduce [Technical Difficulty]. The first [indiscernible] attracted millions of users without its unique and independent product, it's integrated into other products to do specialized agents. Besides what's demonstrated by Irene, under the scenario of e-commerce, the Meitu App, the homepage can also use the RoboNeo. Its original positioning was a productivity tools. After its launch, we found that many customer, end users would use it to do very fun emojis and stickers [Technical Difficulty], the RedNote, some social [Technical Difficulty] demonstrating the productivity scenarios, we can see some creation scenarios.

First, we use it to do brand design. But I guess most of the investors present [Technical Difficulty] design nature. So you may not be familiar with brand design. Before demonstration, let's have a [Technical Difficulty]. It includes 2 parts, basic and application, logo, sound and color. The application is complex. For example, if you are a visual designer, you want to have a brand design for a coffee brand, you can see the filter paper and the coffee bean [Technical Difficulty] color tone, the packages, the advertisement brand menu, the apron and also social media posters. It is based on the style of the basic part. It's just extended. By the way, the previous video was generated by Kaipai within less than 1 hour at home, but I am still a newbie in video making.

So now the example have just been generated by RoboNeo. Now as an agent product, the thinking process is actually quite fascinating. So [Technical Difficulty] I first put it [Technical Difficulty] actually matching the style. So this is the industry know-how because if you're not a major in design, perhaps you may not be aware that the design [Technical Difficulty]. So if the overall style was wrong from the beginning, you would not satisfy the producers' demands. So you can see that the RoboNeo providing 3 different styles [Technical Difficulty] among the images that I like. Then based on my preferences, RoboNeo would make some analysis. For instance, the patterns, the color matching, et cetera. Eventually, based on the analysis RoboNeo would provide solutions -- design solutions. We can see that [Technical Difficulty] and fundamental parts that we just mentioned earlier, for instance, the fonts and the matching colors and logos.

So initially, you can see that the combination of cups and a star and the second is the filter with coffee beans. These are the logos. And the overall matching would be the light brown of coffee beans and [Technical Difficulty] and aside from the fundamental parts, there will also be extensions of application. For instance, the packaging, the design of internal designs of stores, the menus, et cetera. It generates many e-commerce materials for us simultaneously because all of the produced materials are designed based on the styles that it generated at the beginning from the menus to the internal designs of shops, all the colors have used light brown and blue-ish green. So you can see that the produced results by RoboNeo is a complete set of brand designing, not a scattered design of images. So these, to the left, we see some fundamental designs and eventually to the packaging, the [indiscernible] products, et cetera, it has a clear logic from simple to extensions.

And the understanding of RoboNeo [Technical Difficulty] the layers separated, you can see that all the different layers can be moved independent [Technical Difficulty] in this image. For instance, the combination of the filter and the coffee bean and also the worst [Technical Difficulty]. There are many further added functions. Unfortunately, these functions cannot be realized by GenAI at present. But there are many things that RoboNeo already can do. I don't know if you have browsed AI cat and similar themed AI-produced photos on social media. These are becoming increasingly popular fast on social media. These kinds of videos can be made by RoboNeo. For instance, here, firstly, I upload a photo of a cat, and the RoboNeo would transform the style into [Technical Difficulty]. You we can see that the black ears and the white belly have been replicated in a fluffy-doll manner. And now I ask RoboNeo to create a script based on this fluffy-cat doll.

So for 20 minutes, I asked RoboNeo to generate a four-scenario story of a cat at home [Technical Difficulty], et cetera. So we see that the plot have revolved around staying at home. Such kind of a theme is quite popular on social media. Perhaps we've seen topics and videos like these, cats staying at home and enjoying themselves, et cetera. As a follow-up, I asked RoboNeo to produce videos [Technical Difficulty]. You can see that the images of the cat is quite consistent across all the different scenarios. [Technical Difficulty] video sections, I ask it to compile them together and generate a video so we could see that the video I generated is quite stable. So the dynamic images will not produce any flaws that disobey the loss of physics, et cetera. We could see -- and also the images of the cat -- the fluffiness have been well maintained and well retained as well.

And then once [Technical Difficulty] of 20 seconds here. And we could also add a BGM, and this is the eventual result from beginning to end. In fact, this kind of short video can be a complete plot. With some final touches, it can be published. So now we can see some innovative features and functions of users using RoboNeo. So these are the changes brought by AI.

[Technical Difficulty] about photoshops. For instance, if the images -- if we need to photoshop the images, photoshopping our wrinkles and our apps, et cetera, we will not say that we photoshopped them. Sometimes we will say that, in fact, they are our own, but we know deep down that they are photoshopped. So a decade ago or 5 years ago, there are some technological innovations. But now we see the emergence of AI, these AI-powered functions are making these images more convincing and more easy to [Technical Difficulty]. The AI helps users to disseminate these images and videos more conveniently. So that you can see that these kind of functions that are listed on the PPT here are [Technical Difficulty]. For instance, suppose you are a pet owner and you see a video like this of a cat dancing. So you will think that's very cute, and you will try. This would compel [Technical Difficulty] friend.

So we can see there are different templates for dances. You don't even need prompts. You only need to upload a photo, and they will readily available [Technical Difficulty]. Another example for male users that do not like to use photoshop, there would be scramble artist styles for these kind of users [Technical Difficulty] on the platform Douyin, for instance. And also during Chinese Valentine's Day, [Technical Difficulty] some Chinese Valentine Day's videos that interest girlfriends and boyfriends.

All the movements are quite natural. And in fact, there will also be parents using the images of their children to redress them into top learners, for example. And there are also, of course, photoshop functions. You don't need to click on different buttons. You could simply need to put in short prompts [Technical Difficulty] photoshop image, giving you a great vibe. So in short, RoboNeo brings a many -- a creative imagination space for many users. There would be many user-generated content that had been voluntarily demonstrated by different users, and we encourage you to use RoboNeo to explore for yourselves. And the paid services of RoboNeo currently in China, it costs CNY 68 -- or around CNY 68 per month. And overseas, it costs around $15 per month. Now based on different pricing [Technical Difficulty]. So above are the demo of RoboNeo. Above stated are all the demonstrations for today.

Before we go into the second session, we can take a short break. Let's meet later.

[Break]

Now we are moving into the session of the Q&A with the management. Before we start the next session, I would like to thank you again for sparing your time to attend today's Investor Day. I hope today's discussion can bring you some valuable outcomes. Before we move to the next session, I would like to give a brief introduction because we are during the period of the half year results presentation and the final presentation of the year, we will not show more statistics about our running and operation, but you can question about our positioning of the products and the company's overall [Audio Gap] some insights. Thank you for comprehension and support.

[Operator Instructions] About the management present today. Our CEO and Chairman of the company, Mr. Wu Xinhong; our CFO, Mr. L Ngan, Gary; and our Chief Product Manager, Chen Jianyi.

Recently, we saw ChatGPT. We have some possibilities of further integrate models and AI.

ChatGPT may be an app store of the new era. We can see for the first batch, we have cameras. They are integrated into ChatGPT. This is the future direction for application, how to have business cooperation with AI agent and assistant.

In the year 2023 with the rise of AI, what's the boundaries between AI and us to the year 2024 and the first half of 2025?

Actually, it is not a big concern. The empowerment of the AI can have a bigger picture. But now we started discussion, especially when the launch of Banana, we also discussed about with Gary. The progress of the capabilities of the AI and the boundary with us, this will be a pain point for the market.

Can you elucidate on this point?

The management answers, I say first -- every time when we see a drastic increase of the AI capability, the relation between model and application, whether it is synchronization, replicating or replacing or empowerment with the new model Banana is capability in video making. We are relatively positive about it from our own perspective.

Our development with models actually [Audio Gap]. We rank the top in App Store in many European countries. We are digging the latest function of the model. If we apply it to our products, services help to increase its cash ability. For some vertical scenarios. May not be that easy to cater to the needs of the users.

To offer a professional or even industry-level services for users. For X-Design, as an example, we dig into the scenario of e-commerce, we will spread and extend to other vertical scenarios. The services included and needed are not imagined as or can be done by both. We have more open ecology overall. We are relatively positive. With the growth of AI model, we are also growing meanwhile. And we have clear vision about each.

Now as a supplement, suppose 5 to 10 years later, the models become more popular, almost covering all the scenarios we need. But now the models would be a generative, a general product. Its main target audience would be the general users. General users who are not that -- do not have that much expertise, but would like to explore many different areas. We use imaging as an example. We will see that in the Photoshop in designing that has many roles.

We still look at all the visual materials [indiscernible] they experienced designer using Photo [Audio Gap] effect. So when the models become more powerful, there will still be opportunities for platforms like Meitu app, like [indiscernible] et cetera, because that in certain specific scenarios, the general AIs will not be as efficient as Meitu and others that are designated to these specific scenarios because these designated apps would have better results as they're more customized to the different effects. So now if we see even with the general app like Photoshop, there are still detailed [Audio Gap]

So we believe it is the same [Audio Gap]. So there will be 2 questions asked for each person. Thank you for your question. My question is based on competition. I have been a user of Meitu for many years. I now I see Taobao AI can [Audio Gap] but all of these functions in Meitu are paid services. So I would like to ask for users of Meitu, when we do not have that much demand. So I would like to ask for the Banana for Taobao AI, et cetera. Their competition with Meitu, will Meitu change the strategy for about this paid services, et cetera.

For me, this is our emphasis of how to better help our users to [Audio Gap] all the kinds of paid services in Meitu without charging too much. So if you're looking for free device can perform free services relatively well. But the problem is if your product has some -- need some services and is divided [Audio Gap] now in Photoshop, many people have few demands for services. So our new customers while retaining others for paid services. That is our strategic goal. In fact, Taobao or other AI, once they become more popular, because it's been around for some time already.

And from the past to present, our MAU increases, et cetera, have not been incompatible with the growth of AI. I believe in many scenarios in this area, aside from products that can complete the services, the quality and efficiency matters more. For instance, Wink, the quality restoration. In fact, most AI can do quality restoration. But when using Wink, users feel that Wink is the best or at least one of the best in quality restoration. That's what compels users to use Wink instead of General AI. So when you think you need to have the best effect, you would use Meitu and other designated apps. So that is what we think Wink, Meitu and others can provide more in terms of the vertical categorization of products. So I believe at the end of the day, it still depends on the efficiency and productivity.

So my follow-up question, overseas, [Audio Gap] is becoming more popular and ranking first in app stores in many countries. But the Banana in Google apps, it is also becoming popular as well. So I would like to ask, could we predict some of the future users for the AI group photo functions, et cetera? What are their potential demands, et cetera?

I would like to add to the following question. One thing is inevitable is that in the future, as large language models and AI models become more powerful, I believe they can cover some services for Photoshops. This is inevitable, for any user in a real-life scenario. For instance, the inherent and deep level demands of a user are different or at least vary. For instance, they need different makeup.

In many cases, their rationale would be to use the complete set of services provided by Meitu following [Audio Gap] of course, AI can cover many superficial or surface level demands. But for the [Audio Gap], I think AI models pose a great threat to our development strategies. In terms of the AI group photo and the AI functions you just mentioned, the human power predictions cost varies. But sometimes I believe it also depends on the time, the geographical location, et cetera. An example would be, for instance, if we look at [Audio Gap] Software.

The interface tabs, you could see many windows for many templates. In the past, the templates are quite good, but they were not big hits. Whether they're popular or not also depends on whether the users can make full use of them, whether users believe the templates are useful or not. So for sometimes even if it's not as good as others, if it appeals to the demands of a user, it can still become more popular than others. So in the future, we hope [Audio Gap] others, I believe we need to do 2 things. One thing is to ensure that we are capable of marketing ourselves for us to become more popular.

And secondly, we need to have more brand collaborations with content producers. In the past, 3 to 4 years ago, Meitu had a service where you could make video -- short videos based on just a simple image. For instance, you [Audio Gap] do many adjustments with it. For instance, you can even make yourself disappear, for instance, and you could post it to your moments and have interesting interactions. So in many cases, you can't simply have a template and just leave it there. You need to have more content producers to decide on how can you use this specific template. So the AI figures you just said we have similar functions that can make it more popular among whether in China or overseas markets that can help us to popularize these products, including, for instance, the RoboNeo emojis we mentioned earlier.

Recently, I noticed that Meitu large interesting function is photo editing with your friend. In the past, if girls to group photo, they added one by one, but now they're like coordinating work online. What's the methodology for Meitu to grasp and tackle the pain point of users, not like other apps. Maybe in overseas market, we have different products in different regions. How can we find those pain points of different people in different regions? And what -- how can we power them?

What's the goal for mid or long term? The management answers. For functions to develop, we will have [Audio Gap] about the product operation, how do we verify our demo. With upgraded AI capability, this interior intense direction is now engaging more and more customers to co-develop with us. For example, the launch of RoboNeo, we achieved the outsourcing of creators or even designers, they can tell their needs. After we achieve they share it to the social media, we gain reputation.

In turn, it boosted our statistics to tell us we can better enlarge or empower this section. Encourage outside creators and designers to help us better develop our function. You can see from this year on, our capability to catch the hot topic is becoming much more stronger. We are relatively open to the outside creators and designers. This mode brought about many output.

And the answer to another question. About the market share, I didn't check or update very quickly around more than 10% steadily [Audio Gap] all our products. In the past 2 quarters, this figure is an increase. Especially in Euro American market, we are pushing more efforts recently. Basically, about August and September interior, we redesigned our direction and we give more efforts. For example, we can see the top of app store in many European countries. This is also the result of our redirection.

I would like to celebrate the Meitu for this Investor Day. We see that the promotion of Meitu app in the European market, why choose the Meitu?

These are the consumers' choices being chosen by the consumers [Technical Difficulty] other apps may simply [Audio Gap] is designed to serve the entire international community. It has 17 years of [Audio Gap] and it has many services and functions. In the past, these functions were isolated. So you need to have different inlets for these functions to use them independently. But now AI is reinvigorating all of these functions. We said that we could virtually invent -- reinvent Meitu App again through generative AI. So the AI agents can help us better, for instance, connect the different functions of Meitu App to better serve the customized demands.

For me, it is -- for Meitu App, it is a significant advantage. It is an all powerful product in regions where demand for the products are high, for instance, in Europe and also the standards are high, it stands out compared to other products. It means that we can capture the trends and issues of specific regions in a timely manner. And then we also have some products that have been less popular over the long run. And we would, for instance, [indiscernible] in photo editing in editing, et cetera, editing functions as well as Wink, et cetera. So overall, I believe it is the editing [Audio Gap] another thing we should -- and that is all I have to say.

So why -- again, why is it all powerful and has many varieties for Meitu App?

In the photography area, such as the BeautyCam is becoming more popular for consumers as well and also Wink for video editing, et cetera, are also becoming more trendy because a core thing is that product in imaging or video editing areas once it has complete functions, it means that it can deploy AI agents flexibly in almost all processes and to satisfy the consumer demand. And this is a big advantage for any company. And also for cooperation and competitiveness. So again, these are the important points for the status quo.

Next question. I'm an investor, what is our products to see the functions about them? Can we say with many brands or products have functions, but we are more exquisite and more vertical, whether this is our direction or methodology or in the future with -- in the era of AI, we may not see more distinctions, but we focuses on more verticals?

The management answers. First, we divide this part into 2 capabilities. First, about the model capability; and second, the traditional client app or computational design capability. These 2 parts are synchronizing. We cannot solve all needs by using AI. We integrate many other capabilities besides the AI capability, and we let them coordinate. This is more about diversified competition. When we have more similar models, we have services of similar quality. But if we do coordination with other functions, it means we can have greater and better output. This is why we did that for the last 17 years. This actually is what we have done. Besides even if the open source model, our collaboration with [indiscernible] with Alibaba [indiscernible] model, many other companies use this model. We do a lot of enhancement ourselves. Regard? Even if we use the same open source model, there is still a large space for you to do. If we only use a model provided to the comment, we will not have diversified functions.

The second question about -- think about legitimate and Juridical security or functions. For example, the privacy of using some certain photos or videos, -- do we have discussions within our industry? And what's your point of view?

The management answers. We are authorized by many ministries and commissions regarding this part of China, and we strictly obey the restrictions, and we are also the first batch of companies engaging in this program. We also proposed our valuable experience to the government and we are staying close communication within the regulatory department. Overseas, for some major areas and markets, we have strict client regulations. For example, we know the European Union, they have relatively maybe the strict -- mostly strict [Audio Gap] so from the perspective of security, we have more complex considerations.

I have 2 questions. The first is the competition barriers and the commercialization. The first question is that we mentioned earlier, the investors -- on the investor side, the application of AI models, as you just said, that the collaboration with Kumoo and et cetera, have competitive and scale advantages. For instance, our collaboration with Alibaba, et cetera, and other platforms will also be using as Kumoo. So what we're collaborating with Alibaba, what are our advantages?

Using the large model -- larger AI models is our advantage, but it's not that clear once others are also using the same model. So do we need to find [Audio Gap] can have more edge and competition, et cetera. Now in AI, there are many strategic collaborations, et cetera. So eventually, I would like to say if we can have more advantages in terms of our technological edge in the past to present, whether it is our design -- designers or coordination, et cetera, I believe the status of our products is becoming more consolidated and well received.

From an investors' view, we would like to see if we can have more returns with greater competitive advantages. For instance, if once AIs achieve a great breakthrough, we need to have more certainty towards what AI can or cannot do for us.

The management answers. So this is a question that we've been asking ourselves for a long time as well. In fact, before the emergence of GenAI, over the long run, we had similar issues that is the -- for instance, the operation systems, the mobile phone operators, et cetera. Mobile phone operators choosing the imaging area.

In theory, they are eating away our space for revenues and [Audio Gap] when can we actually say that there will be no external competition? I don't think that is practical. So from the past to present, I think the competition has becoming more fierce nonetheless. But our own capabilities, our per user income is also on the rise. So I believe [Audio Gap] gain from competition or our competitiveness remains, I believe this is sustainable for us. So we are open. We are willing to embrace such challenges and tackle them head on or even we believe we need these kind of difficulties to constantly remind us to keep us not to be complacent. I hope that answers your question.

My second question is about commercialization. We see that many products, the cash abilities once in the premium, there would be using credits, the Meitu credits, and there will be higher premium prices. So I'd like to ask in the future [indiscernible] in terms of the premium services of Meitu. Do I need to have multiple layers of premium services? In the future, the high-end premiums, what are the rooms and potentials for this? In the short run, both the mid and long run, do these high-end premiums contribute to much of the revenue? Or will this be a scenario in the future as the tokens consumed by AI increases?

For me, I would like to share a status quo. For instance, in the design studios, the per annual use is around 300 ARPU, the ARPU level. But last year, over the last 2 months or so, some have spent around RMB 20,000 in the in-app purchases of Meitu. So you could see that even under this current ecosystem, there could be in some scenarios where the ARPU rate would be 10x than the average. So we could see that for some products, some scenarios [Audio Gap] the reason is simple because our services help them save [Audio Gap] potentially.

In productivity, we believe at present, we are aware that we are not focusing on the ARPU level aggressively. So there are room for improvements. So we are trying to maintain the status quo, but we're also making adjustments in the future. For the product for leisure, at present, I don't think we are aiming for higher and higher ARPU level.

Many ID photos being uploaded into our apps. We see the consumer habits changing as consumers -- more and more consumers use our apps. So there could be specific programs or services aimed at consumer demands that could cause more or less tokens. But I believe the overall dynamics will remain relatively the same, at least in the short run. The ID photos, I believe, if you have used the ID functions in Meitu, you would see that it has no correlation with the premiums whatsoever. It will actually help you pay less for the same ID photos.

So I believe in the image area, the users' demands are diverse. In the future, at least 5 to 10 years later, many niche yet consistent demands can be further exploited to provide better services for people. The ID images can be a potential. So in the future, there could be similar products with potential. And I would also like to add that in recent years, I've been thinking over the same question as proposed by you.

In the Internet industry, in many scenarios, from a technological perspective, once we look at a product -- another product -- once another product uses the same technology, the previous product risks being replaced, but this issue, in fact, has been persisting in the industries for thousands of years. So from a historical perspective, Lululemon produces the same [Audio Gap] with the same style, Lululemon. So it works the same. If you go to any clothing factory, you can make the same garments, but consumers would choose otherwise, which choose one brand over another.

So this -- we need to always bear in mind of the brand's image, of the brand's influence. I believe this is as a significant component of the cooperation's competitiveness. For instance, if AI is combined with other cooperations that are not specifically aiming at photo imaging, they would not be as good as Meitu [Audio Gap] so what contributes to our brand imaging, helping our brands to become better and better in the industry and having a better reputation and more technological edge. Internally, of course, we [Audio Gap]. That is our own say over the matter.

It's a rare opportunity to have 3 presidents here. I'm a research analyst. We are a fan of Meitu company. We are also very happy to congratulate the anniversary of Meitu Company. We are happy to see you [indiscernible] into a product with high productivity and impression. We are also researching on this kind of company who would like to observe a problem from long run. After communication, I think many of us concerned about the evolution of AI. Last time, we talked about the last mile.

And this is the advantage of Meitu. The influence of Banana may be about storytelling because Meitu has strong capability in terms of aesthetics and engineering, and we have a very strong central unit [Audio Gap] so we can create many popular products and be welcomed by the users. But today, my question is some proposed not only for Meitu company, but also tool like or tool-oriented companies from subscription to complete isolated or single purchasing mechanism or fragmented paying plan because we have seen many overseas companies are doing so. Will this have influence on Meitu company?

I think the subscription may be a mainstream for us because what we offer to customers are not limited to several functions. But we can serve the users for a long run in photo video and design sector. For example, I use a very not proper case. If I subscribe [Audio Gap] one TV series because I am contributing. I hope it can [Audio Gap] in the future or maybe even for myself, my pursuit for later in the future. I think this is the similar ideology for me to no matter it is about products for later or productivity tools. We [Audio Gap] but I also see the room or the space for single purchase it enjoys broader things and scenarios. Maybe in the future, at least for the future 5 years, subscription will be the mainstream. The onetime purchase or a single purchase may increase its proportion in our paying plan.

Next question. Specifically [Audio Gap] Meitu's market share is continually growing, but within domestic area, the ingrowth or revenue is decreasing. What is the mix of the domestic revenue? Does it mean that we reached some peak value? We can see the increase rate is downward while the global increase rate is relatively stable. Can you explain?

The management answers, we didn't observe a downward rate. Is there something wrong with the observation? Can there be this possibility [Audio Gap]

[indiscernible] to September.

Especially for Chinese Mainland, I'm not sure from what statistics you draw the conclusion.

Question, do you mean that the revenue increase rate of the Chinese Mainland will also grow [Audio Gap] regulations? Because I see that many times, so I would like to ask what are the latest paid functions and consumer data about this X-Design. And for X-Design and Taobao, what are the development strategies domestically and also abroad? And what are the potential targets for these 2 apps, both domestically and abroad?

Specific goals will be released by Gary. I would like to say about the product. The X-Design has been developing simultaneously, both domestic and abroad. Not long ago, we've been exploring the offline studio demand. This was mainly because that after spending some time abroad, that we're really interested in the e-commerce materials. But the cost of designing these materials abroad was high, sometimes even hundreds of dollars. This is quite difficult [Audio Gap] for opening a store materials from Meitu, from internal layouts, et cetera.

In China, however, you only need to find a space and find the right personnel, you can make it. But for abroad, the price will be 20 [Audio Gap] so we see this great demand abroad, and we see many that have not yet put their attention on to this demand abroad. So I believe this is currently what the X-Design is doing.

So I would like to emphasize that the e-commerce materials will be one of our main focuses online, but we will also focus on offline studios, helping the -- our clients to reduce their cost of producing materials and helping them to design materials faster and also to have more frequent upgrades in terms of their materials, categorization, et cetera, for instance, coffee shops, the galleries, et cetera.

In the past, often, every upgrade takes around $100 to $200. But now -- and the existing tools cannot satisfy the demand. But now with the assistance of X-Design and AI, we can help them make the situation more convenient. This is a huge opportunity for us. The E-commerce materials can be related to thousands of stores and corporations. [Audio Gap] we can make it much better [Audio Gap] Kaipai.

For Kaipai, a major advantage, both domestically and abroad. Domestically, it is focusing on talking videos, making it a big hit. Before Kaipai, there were many products have similar functions, for instance, adding the subtitles and doing quick edits of the subtitles, but because these [Audio Gap] Kaipai, so Kaipai is focusing on this niche demand [indiscernible] instance. There would be many of such talking videos being produced.

As an example, if we look at the real estate intermediaries when selling houses, they would add a layout of a house with several introductions, but now we see many intermediaries are using talking videos to demonstrate the housing. The same goes for selling insurance. In the past, the insurance promotion videos would resemble typical ads, but now it is different. So we see that many -- the videos that compose mainly of images and words in the past to sell services and now transformed to using videos.

And this Businesses and its competitiveness enhances competitiveness. This is Kaipai's unique advantage in China. Overseas, there is a competitor called [indiscernible], but initially, we did not notice this competitor. But now we're becoming more aware. That overseas producer service provider is also focusing on talking videos. And now I'm excited about this. I'm excited about the potential exchanges and cooperation.

And we're focusing currently on the North American market, et cetera, because it has great potential. So -- that is one way to develop. The second pathway is that coming up with brand-new strategies. Overseas, we have a positive growth rate, and we believe we're trying out new things, but from the long-term perspective, we hope that our subscribers can be more diversified, complement each other, not compete with each other in a malignant way.

The goals, we have not been very clear [Audio Gap] in terms of the development direction, perhaps the X-Design might need personally, I hope that it can raise this market value. But for its independent pricing or in terms of finding more products that can be used by others and more computing points, we believe it is -- we hope more the better.

And for instance, when we're collaborating with [Audio Gap] we hope that in scenarios such as using product image [Audio Gap] whether it is the service providers or the consumers, we have similar demands. We believe we should still make our quality the best of the best.

Due to time limitations, we have the last question.

Last question. I'm Jerry from Hong Kong's Capital [indiscernible] First question. Previously, a question was asked, how to be integrated into overseas agent? The agent of the [indiscernible] and app, what's the schedule and the future plan? And second, strategically, you mentioned in the era of AI brand matters [Audio Gap] for others?

We propose or launch [Audio Gap] model. We still insist on independent development even if we can future [Audio Gap] we purely rely on others or even [indiscernible] by ourselves. Will include operating systems and app store companies and also some companies doing application development. This is an eventual and definite [Audio Gap] you are situated in different eras.

This platform [Audio Gap] everything very possibly. Also integrate all those models or functions and gradually turn them into their own. It's actually not impossible. Many vertical demands, they can never do it. They need a third party, but sometimes they need to do demo or they need to do sample, they can give their own functions or products.

Actually, we are supplementing each other. The agent of the [indiscernible] app, the plan for it answers 2 phases. First phase, we integrate the agent into it. If you use it now, you can discover this is just an entry. Also, you can jump through different functions. This includes some agreements, but second phase, the core client is coordinated editing. And the second is about co-creation or ecology. [Audio Gap]

It can withdraw or it can use agent halfway when using a traditional function. We can be totally layered.

It means, first, you can redo or undo. And second, you can generate formulas, you will combine agents with traditional generators or functions you can share it to others to [ cope ] the ecology, and the future creation of creativity that comes from content makers. We opened the mechanism, it's already doing itself.

In August, we advertised this function because content creators, they achieved some impact, and it went viral on the Internet. So we can combine this with our main product to see some chemistry. The main app is designed mostly by us but I believe in the future, they will develop some creation there own, and they can do the spread on the Internet. This is the co-creation, the second phase.

If we have adequate time, thank you for your support investors. We hope we can have more collaboration in the future, and I need to hurry to another meeting. The corporation or within our company, we did it on our own, the answer. We hope more outside creators can participate in this process. One last question.

My question is about design. The demand side, I am not quite sure my straightforward feeling is that in the past, e-commerce merchants will ask third-party designers to help them solve this kind of materials. But now the 2 business products, what are -- how much of the advanced technological solutions generative AI? For the emerging solutions, the existing actors are also enhancing the competition in the sector.

So what are -- do you think are your potential competitors? How will they influence the market share for instance, from an investor's perspective, I would like to see what are the scale of competition and what are the -- your management answers?

I would like to answer them step by stack. First I hope that we would not think of X-Design as 2 business product. It is not a 2 business product because from our perspective -- from almost any perspective for consumers, the X-Design is a product in between 2B and 2C it helps experts designers as well. So they represent, for instance, content creators for the buyers for all kinds of operators, et cetera. Do the math are that they need to design materials, but they do not have design as their know-how, but they still need to sell their products based on these designs as a type of marketing. So these groups of people, by a large sense, they are not a big cooperation. There's simply a small team of 5 to 10 people, for instance. So before they had to outsource these kinds of services to external parties in the past, it meant to cause hundreds of yuan to outsource them. And the external designers would help them produce these materials.

By now, they have a substitute that is x design. For the market overall, I believe, domestically, the competition is quite fierce because if we look at revenue alone, the similar products, the ARPU level is quite high. So domestically, we are not putting that much attention on the tension because the tension is really quite fierce. But for overseas, it is slightly different because it is an emerging competition. For some products, we have more and more Gen AI start-up teams. So in fact, if we look at the competition landscape, in many scenarios, we need to prepare in advance, prepare for the future.

Not long ago, in the U.S., when we were communicating with Adobe, they proposed an idea that was inspiring saying that the products proposed by Adobe in the 1.0 was a prototype model for the future already. And for camera, these kinds of algorithms, their 2.0 products are already replicating the functions of Adobe. The issue was that the replica, the templates largely relied -- Canva largely relied on the human power and the marginal cost. So if you have Canva when you're editing a menu, for example, or editing a poster, after editing, you will see that there will be some minor flaws or when you're switching to another poster, you would see that they're not that compatible between different templates, different replicas. So this is a conundrum for this 2.0 products.

In the AI era now or the or most existing template will be replaced by AI-generated templates. So in the AI era, as many materials are generated, one thing is that there will be endless creation. There will be endless imagination. So this is one thing we discovered with the agents. Within X-Design once plugging the agent, we see that the demands of users are becoming more diversified.

The conventional limits of designs used to limit the demands of users as well. For instance, in the past, we know that doing the product images, we will simply put them in the middle and add a template. But once AI is added, we can see that we can have many fonts, many adjustments for the image. For instance, we would like a product at the end of the table and sun penetrating the windows and shining on the product.

So this kind of lighting that we can only describe by words, but if it's a scenario that we desperately need, it can only be generated by artificial intelligence. If we would like to do it in reality or generate a template from scratch, it would be extremely difficult. So in the AI era, AI elevates all of our certain points. That is why we're seeing more fierce competition. So I believe this is not only a times but also a major opportunity for designers in the future.

Again, another follow-up question. In the 3.0 or the AI era in the future, scale and magnitude of these kinds of technological deployments because we would like to hear about your estimates, about the market potential.

Numbers aside, firstly, if we would like to imagine the scale of this. In fact, many e-commerce platforms have been trying one thing that is using template. The main issue was not incompatibility was that the shared diversity, for instance, some like a red background or for the selling points of a product, there would be other appealing points that were not present in the template. So these capabilities of an agent or a human designer, but rather whether the producer can generate enough selling points that can satisfy consumers' demands. So we produced this dynamically, automatically and changes according to the scenario, this brings us with great potential as we see today that AI can adapt to different scenarios.

These kinds of opportunities compared to companies without AI are better, more convenient for us to turn into cash. For instance, we see on Amazon. Over the past 20 years, we see more merchants adding color, some people are adding more details to their products on Amazon, for instance, the background images. To generate these background images, you either ask a human designer or use AI for help. In the current era, you can use outsourcing and templates as well. Both can support your production but in the future, if you would like to use ChatGPT or similar AI that where X-Design could have great potential. This is hard to realize. But once we do realize that, the opportunities are great.

Thank you for your participation and questions. Due to time limit, our discussion ends here. That's the end of today's session. Thank you again.

[Operator Instructions] Next, please welcome Mr. Wu Xinhong to share the development of the company.

Thank you all for joining us today. Under the productivity and globalization strategy and subscription-based business model, we have achieved growth in both user scale and profitability during the first half of 2025.

Financially, in the first half, our total revenue is RMB 1.82 million (sic) [ RMB 1.82 billion ], a year-on-year growth of 12.3%, among which our Photo, Video and Design Products segment grew 45.2% year-on-year to RMB 1.35 billion, making it our largest revenue contribution business.

The overall gross profit increased by 27.3% because the operating business is lower than gross profit expansion. So the adjusted net profit attributable to owners of the company reached RMB 470 million, a year-on-year increase of 71.3%.

Operationally, our MAU of June 2025 grew 8.5% to 280 million. Those outside Mainland China reached 98 million, growing even faster at 15.3%. From the users' needs and AI technology, we launched several hot functions. For example, the AI Wardrobe February made BeautyCam top ILIs free charts across Southeast Asia, while AI Flash in June gained popularity across Asia and Europe, reaching approximately 1.9 million users in its first month globally.

At the same time, AI improved our core functions significantly, increased users' willingness to pay. By the end of June this year, our overall subscription rate reached 5.5%.

But we consider AI not only enhance our product capability and monetization, but also unlock a brand-new interactive experience. Transformation in user interaction has often been the force driving technological innovation. For example, the rise of smartphones began with the shift from buttons to touch screens. As AI technologies become more mature, we think it's time to redefine users' interaction.

So on July 14, we launched RoboNeo, our AI agent dedicated to the photo and video industry. It is a multimodal AI system that combines computer vision, natural language processing and generative AI. It aims to be a smart accompany for users. No matter you need a logo, a live poster or an interior design sketch, you just need to type some words, you can get high-quality video or images.

And it even has aesthetic judgment good as designers. It can evaluate its balance audience relevance and other factors and can propose improvements. You can refine outputs from text prompts or reference images or use Meitu existing tools to further fine-tuning.

Although RoboNeo was initially positioned as a productivity tool, many users have already discovered its use in daily life. For example, they turn their daily photos into different memes and spread it on social media.

So without leveraging any of our existing products, the RoboNeo has surpassed 1 million MAUs while topping the category charts in app stores across China, Vietnam and Spain. Looking ahead, RoboNeo will not only be a stand-alone product. We plan to integrate its core function into other products to further empower vertical use cases.

Additionally, I would like to share a few updates on our strategic collaboration with Alibaba. In the era of large models, we will leverage Alibaba's Tongyi multimodal and Wanxiang video model to enhance our vertical-specific training, further refining product capabilities such as the new version of live emojis. They are rich in details and more blended in colors.

Alibaba also provides us with specialized support for training frameworks, data processing and parameter selection, help us to shorten the iteration cycles and deliver better results.

As for e-commerce, we advanced in two key areas. First, we want to launch AI try-on feature combining the Meitu's AI capability and the apparel resources on Alibaba to create a better online shopping experience. Second, we will integrate some DesignKit functions into AI product image into Taobao and Tmall's merchant portal, help sellers to discover and experience our tools and capabilities.

For cloud services, RoboNeo computing needs continue to increase. Alibaba Cloud can provide responsive and simple computing supply to ensure stability. Besides, our model coordination server has been on Alibaba's [ Salient ] platform. Alibaba is also assisting with promoting our image and video editing.

Now I'll give Gary the floor to talk about the detailed business and financial analysis.

Thank you, Xinhong. The first half, our total revenue reached RMB 1.82 billion, a year-on-year growth of 12.3%. It's similarly moderate, but we're initially scaling down our noncore business solutions for beauty industry.

If we look at the biggest contributing segment, PVD products, it reached RMB 1.35 billion, a year-on-year growth of 45.2% basically from the rise of paying subscribers. As of June 30, '25, the total number increased 15.4 million, a year-on-year increase of 42%.

For Products for leisure, the total number is 13.6 million, a year-on-year growth of 39%. With GenAI technologies, we can efficiently and comprehensively tackle pain points that cannot be solved by traditional tools to further increase users willing to subscribe. For example, the AI Flash launched in June, a backlit portrait photos, traditional users may just adjust the brightness or contrast, which could often lead to overexposure. Advanced users may use free tools to enhance based on their understanding of light and shadows, but the process is still time-consuming and requires certain level of artistic background.

Now with AI Flash, we leverage large models understanding of photos to simulate realistic interactions based on photos elements generating authentic effects. So we can see the AI features can help for both beginner users and advanced users to drive their subscription.

For Productivity tools, the paying subscribers is 1.8 million, a year-on-year increase of 69%. Our talking video production, Kaipai, achieved 330% year-on-year growth. This success underscores the effectiveness of our strategy in vertical use cases, and it is now customized to industries like insurance, education and training skin care influencers, so we can help our users more targetedly.

Besides the DesignKit's subscribers reached 56% year-on-year growth. Besides the AI features, our detailed operation also helped our retention.

In the Advertising segment, revenue from this part increased 5% to RMB 434 million, primarily from pragmatic advertisement. We are still cautiously optimistic about this part. As we mentioned in the beginning, we contracted the Solutions for Beauty Industry segment on purpose. So it declined by 89% year-on-year to RMB 30.1 million. We want to focus more resources and energy to our core PVD segment.

From the regional perspective, the part outside Mainland China grew by 35.9%, outpacing Mainland China. The revenue from outside Mainland China accounted for 36.6% of the total revenue, showcasing our progress in globalization initiatives. For the gross profit, total reached RMB 1.34 billion, representing a year-on-year increase of 27.3%. It is 8.7% higher than last year. It is because the change of our revenue mix or structure. The high rate PVD continues to increase while the solutions part continues to contract. So the overall operating expenses reached RMB 290 million, grew at 43.1%. This ratio is 3.5 percentage higher than last year and that is because the Solutions for Beauty Industry is a meaningful revenue. So we can -- so which would continue to decrease because we shift our focus. Hence, I suggest we can analyze the selling and marketing expenses as a percentage of combined revenue from PVD products and Advertising segments only.

So the proportion of selling and marketing expenses was 16.4%, an increase of 1.2 percentage. Among this, the majority of promotional costs are spent on Products for leisure.

Outside Mainland China, followed by Productive tool and Products for leisure in Mainland China, the R&D expenses were RMB 451 million in first half, representing an increase of 6.1%. The relatively low scale increase was primarily due to lower expenses related to foundational model training, a result from two factors. First, we largely completed the training of our large video model. Second, we have adopted a model container strategy, leveraging open-source models, third-party APIs and self-developed vertical models to provide optimal experiences to users. We can avoid dedicating too much resources into foundational model training. If excluding costs related to foundational model training, overall R&D expenses would have increased 14.4% year-on-year due to investment.

In talent, administrative expenses grew by 23.2% to approximately RMB 220 million. This was primarily due to increase in share awards related expenses and additional grants to better align employee incentives. As previously mentioned, the growing contribution of our high gross margin Photo, Video and Design products segment led to expansion in our overall gross profit margin. Meanwhile, operating expenses grew slower than gross margin, further increased our operating leverage and improved overall profitability.

Our adjusted net profit attributable to owners was approximately RMB 470 million, representing year-on-year increase of 71.3%. Taking into account certain noncash and nonoperating items, our IFRS net profit attributable to owners of the company grew by 30.8% year-on-year to approximately RMB 400 million, and our Chairman wants to enlarge the dividend. We used to divide it annually, but this time, we do twice a year. So we will maintain the same ratio, but we do it in the middle of the year, but the ratio is consistent. Next, we will have a special conference discussing this.

Looking ahead, we will stay focused on productivity and globalization and continue to embrace AI technology to deliver more extreme experiences and accelerate our own growth.

That concludes our presentation. We can now begin the Q&A session. Thank you all.

[Operator Instructions] The first question.

I would like to know two questions. First, about the first half. We can see Products for leisure, the AI brought so many users and profits. We recognize it and know the acceleration. We would like to know we would have more strategies for the second half? And for the adjustment for the expectations for the whole year.

Answers. For Products for leisure, it is the base of our users. 17 years ago until the release of Meitu app until today, it contributed consistent profits and revenues. And we also discovered the Productivity tools are phasing out of the market. And our core products like Meitu app, we can accurately touch the targeted users, thus lowering the cost for acquiring customers. So we have a relatively good channel advantage. This is not what we can see in the start-ups. Based on this conclusion, we will continue to invest in Products for leisure. We will not overall transition to Productivity tools.

The second question. We would like to know we collaborated with Alibaba. What is the update? Is it fitting our expectations? And when can we see some significant outcomes?

Our collaboration with Alibaba is basically meeting our expectations. We just mentioned we developed for the last 17 years. Our products grew for long term and they regenerate. Our collaboration still takes a process, Note, we just announced, we can see something. We cannot witness immediately large profit increase or user increase. So teams from two sides may build trust and find the match. I also mentioned in the speech, we collaborated in some foundational level. It is overall positive. Thank you.

And for time, we don't have a clear judgment about the rise of users. We are not so sure about it. Considering our own products, we cannot even give out some expectations very accurately. We can only look into the overall situation. I believe it's good.

Next question.

I'm Lydia from Morgan Stanley. Our collaboration with Alibaba, what's the interaction with Tmall and Taobao? Would Alibaba take a part of the dividend? What about the Productivity tools outside Mainland China? About the MAU and subscription rate, what are our further plans?

Our AI Wardrobe collaboration with Alibaba uses Taobao's advantage in garment sector. We followed the trend -- seasonal trend of Taobao and recommended users with best and most fashion styles and matches. Besides, we also have similar AI Wardrobe function. But this new function can allow users to purchase directly because it's linked to the detailed page of the product. So it can transition some profits.

The dividend or the ratio, we didn't talk about it too much, just a standard portal. In the future, I believe if we have quantity, we will see whether business model can advance.

For outside China, the Productivity tools, we can see. We mainly focus on [ vMix ]. We are focusing on talking videos. The MAU outside Mainland China is higher than that of Mainland China. Our Productivity tools outside China didn't transition a lot of profits. But if we refer to Kaipai in Mainland China, it is quite high. I believe the cashability and the subscription rate outside Mainland China is still very confident.

Next question.

With the generative artificial intelligence progressing, how to define the competition with the generative -- general AI companies and factories?

We collaborate more, match our advantages. Just mentioned by Gary, RoboNeo, it is fundamentally a model container not only to Ali, Tongyi Qianwen and Wanxiang, these basic models. We integrate our vertical trained models and other external models and the APIs from business models.

We can see which works the best, which costs the lowest and to make flexible adjustment. We widely collaborated with these AI companies. We can see it is gaining very good effects. We focus on what we are good at, the polishment for our products and functions. We don't pour resources to something we are not good at, for example, training foundational model.

We also saw the technical companies and applicational companies are different. For some companies focusing on models, they want to enter this industry, but they may be not that suitable. So we're just making up each other's disadvantages.

Next question.

About the cost, you just mentioned our user growth is not definite. We -- our goal is not that accurate. How do we make the budget for R&D?

Our marketing budget remains about 16% of the PVD products because every half year, it will not expand this ratio. For long term, if you tear apart the combination of our products, we have Mainland China's Productivity tools and then outside of Mainland China, their requirement for marketing expense is not the same.

For those in China, we don't need too much additional marketing expenses. Why? I see we can maintain this ratio with higher subscription. We don't need marketing to drive it. It can release some resources for Mainland China's Productivity tools and those outside Chinese Mainland.

The ratio is overall fixed, but the detailed subject may be different. We don't do training for our foundational models. So the left R&D expense goes to talent and also smaller training -- training for smaller models. We just said if we take this part out, it is 14%. For the next 3 years, this ratio may be around 15% of the total revenue.

We don't think it's necessary to increase talent to increase more staff. We focus on how to use AI to create better AI models. For example, in the R&D sector, if we want to make a new effect, 10 people to 20 or 30 people, 2 weeks. But RoboNeo, just 5 minutes to create it, and it is online. You can just publish it, advertise it.

We just equipped our past existing more than 100 capabilities in our apps. It can allow our internal designers or designers external to quickly do something to test. This is revolutionary of the traditional makeup of the company.

Next question.

What is the business progress at RoboNeo? Now its subscription is RMB 12 without time limit just for a day. What is your thought behind it, but not a monthly plan. And it is established -- released for more than a month. What about the MAU and the revenue statistics?

First, because this product is quite a rush. So it is not that perfect. We just made a daily plan. By the end of August or the beginning of September, we will have a complete subscription or single purchase plan. Up to now, the MAU is more than 1 million. The total revenue, we are on the initial phase. So we didn't calculate this part separately. The overall outcome is very cheering up because we don't use any existing metrics to share traffic and it is of zero budget initially, but it gained 1 million users now.

Next question.

Congratulations to strong performance. About video generating, how can we regard the future space for it and how to integrate with our existing tools? Will you have some plans for products in this regard?

The video generating, we apply it in many Meitu products. For example, Wink, Enhancer and Portrait Beautification. After that, behind it is generative AI.

Just Kaipai mentioned by Gary also deploys models to do it. RoboNeo, many users use it to do video generating in Brazil. In Japan, they use it to make video. Our MOKI and other products. So the video and design are two aspects we will focus very much more on because it is based under the vertical PVD sector, but we will focus more on these two products. Why?

Our input into video grew faster. It ranked second in China's market. It is faster than the picture products. The cashability is still progressing, so we can do some applications. Our input in the video products can also be used by image products. So indirectly, imaging products benefit from this part.

Question.

What is the model for video generating? It's from the third party or self-developed?

Our foundational model are from external. Basically accurately speaking, it is Tongyi Wanxiang by Alibaba. It is more about video and imaging. For Qianwen, it is multimodal and language. Based on this, we have many vertical models. We train very many of them because we don't need to invest a large number, but they can perform very well and efficiently.

Besides these vertical models, we train expert models from the perspective of -- for example, Kaipai as a talking video tool, can it be a director to help users with train in this part. In the future, our Productivity tools are likely to be equipped with an expert model. They are beyond most of us, our imagination.

Next question.

About markets outside Mainland China. The AI Flash is beyond the expectations of the market. What are our R&D of the new tools or products? Can it refer to the previous successful ones because some are very vertical and very good performing. So do we refer to existing products? And do we have methodologies or what plans do we have to promote better products?

For outside Mainland China, we have Products for leisure and Productivity tools. For the leisure part, we are more -- we are closer to social media. We fit the need to make content or some underlying hot topics. This is the mainstream.

We market with KOL as the base of our marketing. If we can grasp the needs very well and catch the hashtags and we can see very good performance, we will continue to expand investments in this regard, including further cooperation and increase the volume.

But for Productivity, it is not fancy. We will seize very, very vertical scenarios, and pain point. So for the Productivity tools outside Mainland China, we retain the scale, not like Meitu app. It is equipped with many functions. But for outside Mainland China Productivity tools, we want to make it small but targeted more vertical. I can share a previous one.

For those shops, we provide tools for them to design the whole viral effects and designs. Maybe we don't have many shops. But if we look at globally, the small shops can transition to small companies. They can do it offline, but also online. Many needs are connected with each other.

An online question from Tencent Meeting.

Still about the market outside Mainland China. The Kaipai performs well. What about the difference? Is there a difference between the in-China or Outside China? For example, the division, the European, American and East Asian and different divisions, do we have different plans for Kaipai?

First, answers. We focus mainly to European and American markets. We have a new brand called Vmake. We do some distinctions. For those inside Mainland China, they use mobile. But for those outside Mainland China, they use it on the web, web browser. So we can see, overseas users focus more on productivity because they use it on computers.

For users of different regions, American users, they have very high requirements for subtitles. Not only it should be synced with what we really see, it should be quick and smooth. And they also have requirements for minimal effects of subtitles. So it's more professional.

But for Asian users, still for subtitles, they may not require the real time with more stylistic requirements, but not unprofessional. I think this is caused by cultural difference as well. For African part, for Euro-American markets, they don't have much need for it. They don't do too fancy effects. But in Asia, they will use some effects from the shows, TV shows alike.

Asian users underscores the first several seconds of the review. They will do very explosive statement or something. But for Euro-American users, they don't have to catch eyes for the first few seconds.

Next question from Tencent Meeting.

About the subscription rate. What about the 5.5%? What is 2B or 2C overseas or Mainland China? Can we see more details, which part is beyond expectations for the annual payment rate? What is the picture? For the first half, it's slightly decreased than that of last year. What is the future trend or what will we view it now?

We reviewed this part in the report. For the Products for leisure, it is around 5.3%. For Productivity, it is 8%. Based on the figure, I think it is stable. It is the question of the mixed business functions. The leisure part accounts for a big part in China. So it will decrease. So it's just about the question of the mix of our combination of our business or functions, the ARPU value.

Next question from Zoom. Karen from [indiscernible].

It has become apparent in the West that chat LLM apps such as ChatGPT are going to imminently monetize various forms of closed-loop e-commerce either directly via transaction commissions or via advertising. Meitu seems to be in an interesting position to serve as an enabler of that in China through video and photo, asset creation and management in a similar way that perhaps Shopify has done in the West.

So if one were to believe that search-based e-commerce were to fragment in China, my question is, is Meitu open and willing to work with other e-commerce platforms should that happen? Is DesignKit aiming for a much broader set of merchants than just Alibaba merchants in the event that there is a fragmentation of search-based e-commerce perhaps to the LLM apps in China?

So I think the answer is yes. We are actually open to work with any e-commerce platforms. In fact, within the cooperation agreements with Alibaba, we have very specifically agreed that this is not going to be an exclusive cooperation.

So if you look at the number, monthly active users DesignKit have in China is tens of millions users. I think the Alibaba's estimate to Taobao and Tmall sellers is only 6 million to 7 million. We are already serving sellers from outside of Alibaba, Taobao's and Tmall.

I think to your question relating to how we will work with these e-commerce platforms? We don't have a current plan to do that, what you are suggesting, which is very interesting. We'll keep our options open. But as long as it's something that is related to photo, video, that would be something we would want to watch. We do set our boundaries within the imaging and video sections.

Next question from Tencent Meeting.

From the perspective of investors, I bought Meitu, Inc. for several years. The highest market value can reach around more than RMB 100 billion. So are you confident for a short term? Or what's your plan? And what do you think for market value?

Answers. We never evaluated our market value and gave instructions based on it. But it doesn't mean we are not confident in our functions and business. Our philosophy is to make our products good, to build world-class products, to make the product capability, drive the growth of users and revenue and profit. So all the stock price or market value is just a sign of we do our basic functions well. I cannot say -- we cannot reveal too much about it. But thank you for your support.

Due to time limit, the last one question.

I would like to know from AI-oriented industry for overseas models, no matter large or small, they are generating at a faster pace. It is submissive but we see fierce competition. If we see from the industry barriers, for example, the data insight or community operation, these regards, they are core KPIs or do you think it is more about win or loss. We see some vertical scenarios, which are just coming and then the tier, second tier, third may be difficult to get in then test?

Answers. We have competitions inside our company and not inside our company, but inside China. We also need to think for our overseas competitors. The generative AI products, what opportunities or challenges do they brought -- do they bring to Meitu?

We also see the traditional workflow or even organizational structure may be challenged. So we do some revolutions inside. We hope to be like other AI start-ups. We hope to be very streamlined and agile, high efficiency in human resources. Everyone in the team can be versatile. This is our organizational adjustment to make the organization more flat, reduce reports or authorization. We can make the whole process for decision-making faster or smoother. We hope to act fast and be more compatible.

We canceled the Meitu Image Festival held every year previously because this is not then functional -- core functional. So we want to grasp the opportunity like AI agent.

For competitiveness, I would like to mention some points. First, we just mentioned Meitu developed for 17 years. We constructed the 280 million MAU. This is a very strong market channel advantage. Our cost for acquiring customers is much lower than our counterparts. Besides our technical power, we know the competition of AI agent is fundamentally the engineering competition. So if you have scaled your investment into engineering, I believe you are competitive. That's why we launched new projects and products so fast because we made preparations for the last few years. We deploy our core functions into the MTP server for external and internal use and deliver. This is the engineered scaling.

Next aspect, our insight into aesthetics. This may seem like emotional or ethereal, but I think it counts, it matters. So our slogan is to unite art and technology elegantly. We will transition it into the competitiveness of our apps.

In vertical scenarios, it is not about the order or rank. We need to really know the pain points of the users inside this vertical channel. No matter UI or experience, what's most satisfactory. The generative AI may not do their business this way. They want to find the biggest common points inside their users, but some valuable vertical scenarios are missed this way.

So a interesting point is that if we do something that is not targeted, any common user, they can say, some model can do, ChatGPT can do or some other models can do. But if you really ask those who need, they will say, Meitu does better.

It is to some extent, they rank the order, but it is not because the order you came, it is because you know the pain points.

Next question.

Gary mentioned previously, nonexclusive cooperation with Alibaba. We also heard our AI applications are based on Ali's Qianwen and Wanxiang model. So why among all these open-source models like DeepSeek and closed-source, why we chose Ali's models for applications? What's the reason behind it? Is there an agreement?

Answers. If we see the big picture about the competition of AI, first, we must exclude those closed-source models. We just do API. We don't have many open-source models. For DeepSeek, very popular during the Chinese Spring Festival compared with Ali's Qianwen, Wanxiang, DeepSeek doesn't do multimodal. It can generate video or image.

So for us, it's basically impossible for us to use DeepSeek because verbal interactions, multimodal comprehension and image creation, they are not separate. They are a whole. Like RoboNeo, it interacts and it do something those can do. So we need a model to synthesize. Besides the scale for investment, Ali pays more than RMB 10 billion annually. This is a rare thing in open-source model and it tops very often in the charts.

So despite -- so let alone we have or don't have an agreement on this part, it's an ideal option, but we can cooperate deeper. We can do detailed and specialized training. They can open-source the latest version to us because we know the open-source version is usually later than their advanced ones. So we are more competitive.

About the concept of model container, but we also contain other models outside of Alibaba. For different effects, we may use different models. So it's basically about how we understand which model fits the best scenario. We just introduced it into our own model container and we let them compete. The best ones will be selected and used by users. We also gained valuable experience from this process.

Thanks for questions from investors and analysts. If you want to know more about the company's results, you can contact IR team. The presentation for investors ends here. We will not provide further simultaneous interpreting. Thank you for your support. Hope you can focus and support Meitu in the future.

[Statements in English on this transcript were spoken by an interpreter present on the live call.]