Rather than competing in general capabilities and focusing on the device side, Lin Songtao, Vice President of Tencent, noted that Marvis is dedicated to excelling at system-level operations.

Article by WANG Xinyi

Edited by ZHANG Yuxin

"Technological breakthroughs determine how fast AI can advance, while the ability to deliver real value determines how far AI can go." At this year's WAIC Tencent AI Application Innovation Forum, LIN Songtao, Vice President of Tencent, shared this perspective.

As products launched after the "Claw craze," Tencent's three Agent products — WorkBuddy, QClaw, and Marvis — have embarked on distinct development paths.

First, regarding WorkBuddy, LIN Songtao publicly stated at the forum that WorkBuddy's DAU has steadily ranked first among domestic productivity intelligent agent products. For QClaw, its related business and part of the team have recently undergone restructuring and been integrated into the department behind WorkBuddy, though the QClaw product will continue operating. Marvis, the last of the three to launch, has taken a differentiated system-level Agent route. LIN Songtao revealed that Marvis has gained strong traction among users: its DAU exceeded 300,000 just two days after launch, with a 7-day user retention rate of around 54%.

While WorkBuddy and QClaw remain high-profile topics, we would like to focus our discussion on the Marvis product.

Unlike Claw-style products, Marvis does not follow a general-purpose Agent path, but prioritizes system-level operations. CAI Jiantao, head of Marvis, revealed that in terms of user scenario distribution, local file-related capabilities account for the highest proportion (44%), followed by computer hardware management (28%), browser tasks (18%), and application-related capabilities (16%). Surprisingly, only about 6% of users use it for information searching. "Users have a very clear understanding of the product's positioning, and almost no one regards it as a substitute for Doubao or Yuanbao."

Marvis was initiated as an internal project earlier in 2025, codenamed "Device AI Assistant."

From project initiation to official release, Marvis's design plan went through three full rounds of overhaul and reconstruction. With simple lines, a visual and fun scene of ponies working, and a red scarf that echoes the iconic QQ penguin, it is clear that the team did not want to create an Agent product with a high barrier to entry for users.

As an Agent product developed by Tencent's AppGallery team, Marvis has carried the DNA of an app store from day one. In the AI era, simply continuing the traditional app store distribution model is no longer sufficient — it is critical to focus on whether results can be delivered according to user needs after distribution. LIN Songtao, Vice President of Tencent, told us that Marvis's North Star Metric is not DAU, session count, or other activity-level indicators, but the number of real tasks completed by users.

At the same time, another key concept is deeply embedded in the AppGallery team's DNA: cross-platform functionality.

At this year's WAIC, AI-powered agent smartphones were undoubtedly one of the most attractive product categories for visitors. While edge AI on mobile devices has become a hot topic, CAI Jiantao stated that the team is not developing for the mobile edge.

The reason is simple: mobile phones have limited memory and battery capacity, making it difficult to deploy large-parameter models that can handle complex tasks. Therefore, the edge side they target covers devices such as PCs, Mini PCs, and AI Boxes. Users can use their mobile phones as a control terminal to send instructions to PCs or cloud containers for execution.

Currently, Marvis is collaborating with the Hunyuan team to plan the launch of an edge-side model cluster.

Marvis is not the only AI product from the AppGallery team. Concurrently, the team has also launched a mobile Vibe Coding tool called Toast.

Traditionally, AppGallery's strengths lie in helping apps launch, promote, and achieve commercialization. Toast is designed to create new incremental value on the supply side, enabling more demands to be fulfilled. The two paths represented by Toast and Marvis — one for creating applications, the other for delivering direct distribution of services — may converge at some point in the future.

From the app store to Marvis and Toast, the AppGallery team is trying to figure out the ultimate form of app stores in the AI era. No single product can meet all user needs, and what the team can do is gradually explore the ecological niche that app stores will occupy in the AI era — a question that does not yet have a definitive answer.

During WAIC, we had an exchange with Tencent's Marvis team. Below is the dialogue between media including *Intelligent Emergence* and LIN Songtao, Vice President of Tencent, and CAI Jiantao, Head of Cross-Device Business and Marvis at Tencent AppGallery, which has been slightly edited and condensed:  

More, Better, Faster, Lower Cost

Q: What are the user performance metrics for Marvis after its launch?

CAI Jiantao: Marvis went live at 11:58 PM on May 20. By May 22, its DAU had already exceeded 300,000, with a 7-day user retention rate of around 54%. Since launch, DAU has been steadily rising with a very positive trend. However, at this stage, we are not overly pursuing DAU growth, but focusing more on continuously polishing product capabilities and enhancing user value.

Q: What specific directions are you focusing on for product refinement?

CAI Jiantao: We are pursuing high-value, meaningful growth.

Most Agent products on the market face many pain points, such as inability to access default browser data. There is significant room for optimization in areas including browser tasks, shared Cookies, and permission definition, and we have made substantial investments in these aspects.

For example, in the PC troubleshooting scenario: one day you suddenly notice a new program launching on startup. If you ask Marvis about it, it will check the installation logs and inform you that the program was installed silently alongside another application — a problem that traditional software can hardly diagnose the exact cause of.

Q: What is your ideal user persona for Marvis?

CAI Jiantao: We hope that everyone who uses a device, regardless of the type of device, can have Marvis to enhance their device usage experience and efficiency — covering productivity, daily file management, app operations, and gaming experiences.

We want Marvis to be universal and general-purpose, closely integrated with daily life and entertainment, beginner-friendly, and ready to use right out of the box.

Q: Based on actual user usage, what are its main application scenarios?

CAI Jiantao: Almost no users regard Marvis as a substitute for Doubao or Yuanbao. In terms of user scenario distribution, 44% use local file-related capabilities, 28% use PC and hardware management capabilities, 16% use application-related capabilities, 18% use browser-related tasks, while the proportion of users using it for search is very small, around 6%.

We have also made initial attempts in the gaming sector. When Tencent's new game *Out of Control Evolution* was first launched, we collaborated on game-related features including map exploration and strategy acquisition, which received excellent feedback and achieved strong penetration rates.

Q: You emphasize Marvis's system-level operations, but general-purpose Agents today can also perform many system-level tasks. What makes Marvis different from these general Agents?

CAI Jiantao: The four core elements of an Agent are perception, thinking, planning, and action — starting with perception.

Perception means the Agent can proactively understand devices, users, and the environment. With user authorization, it knows what files are stored on the device, what each file represents, what programs are installed, and even past device behaviors, browser activity, and transaction authorizations. Only by achieving system-level integration can this level of perception be realized.

LIN Songtao: Being able to perform a task and performing it well are two completely different things. The reason we built a system-level Agent is to improve task execution efficiency and the quality of task completion. The AppGallery team has over a decade of experience in cross-platform and system-level work, focusing on the "App x System x Device" ecosystem.

Many current Agents complete local file recognition or processing through visual methods: screenshots are uploaded to the cloud for recognition before results are returned, which is inefficient and far from optimal in terms of time cost. Actual tests show that performing these tasks locally in conjunction with the operating system delivers far higher task completion rates and accuracy than current common AI solutions.

For example, the problem "Why is my computer running slow?" has existed for 30 years. What has always been lacking is not knowledge, but a deep understanding of that specific device.

Marvis does not fully rely on cloud large model transmission and judgment. Instead, it leverages system-level capabilities to maximize local performance, then achieves edge-cloud integration. A more, better, faster, and lower-cost approach is the real solution for the future.

Q: What is the relationship between Marvis and system vendors?

LIN Songtao: I think it can be viewed from two perspectives.

First, our relationship with system vendors is more complementary than competitive. We have long-term partnerships with companies like Microsoft and Intel, and Tencent AppGallery for PC is a core component of the official app store ecosystem on Windows. The same applies to Apple: Apple itself develops app solutions for Mac and works on EXE compatibility. However, users still prefer to use Windows laptops for gaming, and our Mac App Store can help many games run smoothly on macOS. Vendors are actually very welcoming of our efforts to optimize app layer integration for them.

Second, from the product itself, Marvis has been cross-platform from day one. Users do not care how a problem is solved — they only care about the final delivered result. For example, users can ask AI to collect information on a PC, edit a document, and then send it directly to Marvis on their mobile phone. Such cross-platform capabilities cannot be achieved by a single-ecosystem vendor. In the future, we will continue to leverage our cross-platform and edge-cloud integration capabilities, which are key service advantages for us.

Mobile Phones Are Not an Ideal Edge-Side Scenario

Q: Marvis focuses on system-level capabilities. Compared with other intelligent agent products, what is its most prominent differentiating feature?

LIN Songtao: The way to judge whether an Agent direction is correct is simple: see if you feel happy or worried when large models are upgraded. If a large model upgrade renders many of your Agent functions obsolete, that is clearly not a direction worth investing in.

Marvis prioritizes the edge side, with its biggest advantages being time cost and efficiency — benefits that cannot be achieved by relying solely on large models. The cloud large model is more like the human brain, which needs to become increasingly intelligent. What we focus on is more like the cerebellum and reflex nerves, ensuring speed, execution efficiency, and stability. Under this premise, edge-cloud integration will yield twice the result with half the effort.

CAI Jiantao: Marvis is an edge-side product that runs locally. After perceiving the full context, it can execute tasks more proactively — it can identify in advance what each image is about, what each document contains, and which files are invoices, schedules, or official receipts. Other general-purpose Agent products, such as Codex, even with permissions, can only perform blind full-disk scans. If they miss the target data, they cannot complete the task.

We have full confidence in Marvis's local file processing speed and accuracy. Since it does not require cloud interaction, all file analysis and processing are done locally. In terms of speed, Marvis is far faster than Codex or other competing products.

Q: Does this mean that Marvis will become a standard feature on every PC in the future?

LIN Songtao: Many PC vendors are already in discussions with us about pre-installing Marvis. We hope Marvis will act as a safeguard that guarantees the minimum level of capability and experience for future PCs, rather than defining their maximum potential.

Q: Which third-party vendors will you collaborate with on edge-side development?

CAI Jiantao: First of all, we will most likely not collaborate with mobile phone manufacturers.

We believe the edge side is more suitable for scenarios such as PCs, Mini PCs, and AI Boxes. We will continue to invest in edge-side model clusters for these scenarios, including small translation models, 1-2B parameter multimodal visual models, and small-size text understanding models, as well as 20-30B parameter models for AI Box use cases — different scenarios require models of different sizes.

Q: But users want Agent capabilities to be more convenient, since they cannot carry a PC with them everywhere.

CAI Jiantao: All you need is a computing container, with your mobile phone acting as a control terminal. This container can be a PC, Mini Box, or other hardware. At the same time, we also provide a cloud-based Marvis service.

The cloud-based Marvis capabilities are a long-term accumulated advantage of ours. Many users do not use PCs regularly, but as long as there is a computing container or cloud storage available, that is sufficient. We are committed to developing smaller-size models, such as 3B to 30B parameter models, which can be deployed in the cloud or elsewhere to become exclusive services for users.

Our existing achievements include that, through our collaboration with Hunyuan and Intel, we can run a 3B model smoothly on Windows with less than 8GB of VRAM, and achieve excellent performance on Mac with 32GB of RAM.

Q: What is the current activity level of the Skills Square?

CAI Jiantao: User penetration is around 37%, which is not particularly high. AppGallery's historical strength has always been operating an app store, which is essentially building an ecosystem. We hope developers can expand their products on Marvis, and in the future we may launch a channel where developers can provide high-quality services to users, who can then purchase these services.

Q: Have you introduced any incentive mechanisms to promote the third-party developer ecosystem for the Skills Square?

CAI Jiantao: We currently have two incentive measures that have not yet launched, but they will be released very soon.

One type allows professional developers to create paid content on the platform, and they can earn corresponding revenue after users make purchases. The other type rewards regular users who upload content with Token incentives, platform operational resources, and even other partnership opportunities.

Building the Next-Generation App Store for the AI Era

Q: Why did the AppGallery team choose to develop Marvis?

CAI Jiantao: Essentially, the transition from AppGallery to Marvis is homogeneous — both are solving service delivery problems, but the process has been shortened from two steps to one.

Traditional app stores focus on distribution. For example, if a user wants to perform device diagnostics and memory cleanup, they need to first download the corresponding software from the app store before they can run the cleanup. Marvis represents an evolution of that traditional distribution model: developers used to create foreground apps, but now they become background service providers that deliver capabilities through CLI, MCP, SDK, or other frameworks.

LIN Songtao: For over a decade, AppGallery has been dedicated to one core mission: delivering digital capabilities and content to users. In the PC era, we distributed software; in the mobile era, we distributed apps. But the AI era is different.

In the past, when we focused on distribution, we did not need to worry about what happened after users downloaded an app. Today, simply delivering capabilities to users is no longer enough — we must also guarantee the delivery of results. The future form of app stores will definitely continue to evolve, which is why our North Star Metric is not DAU or session count, but the number of real tasks completed by users.

Q: What stage has Marvis's commercialization reached? What are your future goals and plans?

LIN Songtao: This is not the right stage to pursue monetization, as the domestic ecosystem where users pay for Tokens has not yet fully matured.

Marvis provides a certain amount of free Token quota every day. The real purpose of this quota is to encourage users to try out real tasks. This quota may change in the future, and we may also collaborate with B-end vendors. Our previous partners, such as Intel and Microsoft, have all expressed strong interest in collaborating with Marvis, including exploring opportunities in overseas markets.

At the same time, the overseas Token ecosystem and users' willingness to pay for tools are more mature than in China, which is a key direction for Marvis's future development.

Q: Some users have reported that when executing tasks, Marvis first installs AppGallery. Does this mean that it is still in the stage where AI calls apps to complete tasks, rather than eliminating the app layer entirely?

CAI Jiantao: Let me clarify first: what Marvis installs is not AppGallery, but a mobile application engine. There are two reasons for this: first, Marvis needs to manage app installation, versions, and updates