WeChat AI Agent in Final Testing, Could Reshape User-Service Interaction

An AI product banned by its own platform ultimately forced Tencent to play its biggest card.

On June 2, Tencent Holdings (HK:00700) surged more than 10% in a single day, adding approximately HK$415.8 billion (about RMB 360 billion) to its market capitalization—the largest single-day gain since January 2021.

The curve is driven by an unannounced rumor: reports suggest Tencent is finalizing testing of an AI agent integrated into WeChat, with plans to initiate compliance review procedures as early as this month. Tencent has not commented, but Citigroup has maintained a “Buy” rating with a target price of HK$763. On the same day, the Hang Seng Tech Index rose 4.72%, and Meituan surged over 9%. A single rumor ignited the entire Hong Kong tech sector.

It has been difficult to recall the last time WeChat’s product developments triggered such a significant market reaction. With 1.4 billion monthly active users, any adjustment from WeChat is major—but this time, the capital market’s valuation logic goes beyond that. The narrative being priced in is that WeChat is transitioning from a platform where users actively seek services to a system where AI completes tasks on behalf of users.

According to insiders, the interaction entry point is planned to be a right swipe on the WeChat home screen to open the AI chat window, without entering any standalone app.

The scope of WeChat's AI agent extends beyond WeChat itself. It is reported that WeChat is collaborating with smartphone manufacturers such as Huawei, Honor, Xiaomi, OPPO, and vivo to introduce Agent-to-Agent (A2A) capabilities. Users will be able to initiate WeChat audio or video calls or send messages to contacts directly through their phone’s built-in voice assistant. Currently, this functionality has been rolled out on select Honor devices: by updating both YOYO and WeChat to their latest versions, users can simply activate YOYO with a voice command to perform actions such as “Send a WeChat message to Zhang San” or “Call Li Si via WeChat voice.”

Its purpose goes beyond being just a chatbot within a messaging window. By integrating a system-level voice interface on mobile devices with WeChat’s communication and mini-program capabilities, it condenses multiple steps—“Open WeChat → Find contact → Type or call”—into a single command.

Behind the June 2 announcement was a year-and-a-half-long strategic struggle. Tencent wandered down wrong paths in grappling with the question of whether AI should be integrated into WeChat, until an unexpected event finally pointed it in the right direction.

In May 2024, Tencent Yuanbao was launched as a standalone app powered by the Hunyuan large model, following a standard consumer-facing chatbot model—users speak, and it responds. By early 2026, Yuanbao’s standalone app had just surpassed 40 million monthly active users, while DouBao had already exceeded 300 million.

During the 2026 Spring Festival, Tencent launched a major social viral campaign on Yuanbao with significant investment. The core mechanism was “Visit Yuanbao to split 1 billion yuan in cash red packets,” encouraging users to share Yuanbao links within WeChat groups to acquire new users through social networks. For several days, the red packet links flooded friends’ circles and group chats.

On February 4, the day of the first withdrawal of Yuanbao’s Lunar New Year red packet campaign, problems arose. The official WeChat account “WeChat Team” issued a notice: upon receiving user complaints that Yuanbao’s Lunar New Year event induced sharing, harassed users, and disrupted ecosystem order, it had accordingly restricted direct access to its links within WeChat. WeChat’s PR director, Zhang Jun, added in a Moments post: “User experience comes first—we treat everyone equally,” accompanied by a meme image of someone hitting themselves while going crazy. Yuanbao was forced to switch to password-based red packets, requiring users to copy a password and switch back to WeChat to paste it, completely breaking the viral chain.

The takedown order came from WeChat, and the target was Yuanbao. Both are under Tencent, meaning that while Yuanbao’s planned budget of one billion is still being spent, its distribution channels have been blocked by its own company.

Externally, two interpretations were read: on the surface, WeChat was following the rules and not giving preferential treatment just because it was its own app. More notably, beneath the surface, Tencent internally had not yet reached a consensus on “where AI should be placed.” Zhang Xiaolong’s WeChat did not allow any external AI application—even those from within Tencent—to access its social relationship chain.

The ban has boiled the issue down to a simple question: Tencent’s AI shouldn’t be developed as a separate app outside WeChat—it’s better suited to be integrated directly within WeChat.

In the following months, a series of preparatory moves were swiftly implemented. In March 2026, Tencent disbanded its AI Lab, which had existed for a decade, integrating its personnel into the HunYuan system to consolidate efforts on foundational models. In April, HunYuan 3.0 was released with a total of 295 billion parameters, adopting an efficiency-focused strategy of “sufficient and cost-effective” performance rather than pursuing a trillion-parameter arms race. Meanwhile, Yuanbao quietly entered the WeChat chat interface in the form of a “Red Envelope Cover Assistant,” no longer requiring a separate downloadable app but instead appearing as a contact within the WeChat chat window—transitioning from an external competitor to an internal component. By June 2, reports emerged that WeChat’s native AI agent was nearing the end of its testing phase, finally bringing the initiative to light.

There is a fundamental difference between Tencent's existing AI product line and the WeChat AI Agent.

Tencent's AI product lineup is robust, with the HunYuan large model serving as the underlying engine, Yuanbao as the consumer-facing conversational interface, CodeBuddy and WorkBuddy targeting developers and enterprises as well as end users, and QClaw combined with the OS-level assistant Marvis covering additional use cases.

However, these products share a common limitation: while they excel at information generation and content creation within their respective scenarios, their ability to coordinate actions across ecosystems is weaker. Hunyuan generates text, Yuanbao provides image and text analysis and recommendations, and they can handle tasks like searching for information or downloading files, but the workflow is still incomplete when it comes to invoking external services or executing transactions across applications.

Meanwhile, competitors have made different progress between saying and doing.

According to QuestMobile data from March 2026, ByteDance’s Doubao has 345 million monthly active users, capable of simulating user interactions by identifying screen elements to complete online purchases. Alibaba’s Qwen has approximately 166 million monthly active users, deeply integrated with e-commerce, mapping, travel, and Ant Pay, allowing users to directly book flights and reserve hotels through it. Yuanbao’s product capabilities are clearly lagging, and as a standalone app, it must compete head-on with Doubao and Qwen on customer acquisition costs. With such a large gap in monthly active users, directly catching up may not be realistic. Tencent’s path forward may not lie in this standalone app赛道.

The WeChat AI agent is designed to bridge this gap—it is not a WeChat version of Yuanbao, but a reimagined product. According to insiders, the interaction interface is planned to be accessed by swiping right on the WeChat main screen to open a chat window. Users issue commands in natural language; the AI breaks them down into subtasks and automatically invokes the corresponding mini-programs to search, compare prices, place orders, and make payments—all within a closed loop on WeChat.

Compared to Tencent’s existing AI products, if the recently disclosed WeChat AI Agent functionality is accurate, it actually fills three key gaps in Tencent’s AI capabilities.

From output to execution. Yuanbao focuses on content generation and conversational responses. If the currently disclosed WeChat AI assistant features are accurate, the WeChat AI Agent will be able to directly perform real-world tasks such as making appointments, ordering food, purchasing tickets, and paying bills—completing the final link from intent to transaction with WeChat Pay.

From acquiring new users to activating existing ones. Yuanbao relies on users to download a brand-new app from an app store, while WeChat’s AI Agent naturally has access to over 1.4 billion existing users—no download, no registration, and no education costs required. This represents the largest distribution potential in China’s mobile internet today.

From screen simulation to API calls. Doubao’s approach simulates human actions by identifying the positions of buttons on the screen via GUI recognition—a method that faces increasing blocking risks from some app vendors and consumes substantial computational resources, estimated at tens of thousands of tokens per operation. WeChat’s approach is different: its millions of mini-programs are essentially a standardized set of structured APIs, allowing the Agent to call them directly, achieving efficiency improvements by several orders of magnitude. Combined with a hybrid architecture that processes sensitive data on the device and handles large model inference in the cloud, all data remains entirely within WeChat’s security boundary.

Liu Chi-ping, President of Tencent, provided an official statement during the Q1 earnings call: "Beyond foundational large models, AI agents with autonomous execution capabilities have demonstrated breakthrough application value. The WeChat platform inherently possesses multiple advantages for hosting AI agents." WeChat’s ecosystem spans communications, social interaction, content, commerce, and payments, forming the essential capability framework for an ideal AI assistant.

The time window is also narrowing.

At the May shareholders' meeting, Ma Huateng offered an unusually candid self-assessment: "A year ago, we thought we had boarded the ship, but later we realized it was leaking. Now we feel we’ve managed to stand on it, but we still can’t sit down comfortably—we still hope the ship can move faster."

Three sentences encapsulate Tencent’s three-year journey in AI. Early bets on the HunYuan model were correct, but execution was slow. Directly competing with DouBao and Qwen using YuanBao fell short of expectations. The rise of DeepSeek added further pressure: even small teams can build top-tier models, and big companies’ dominance is not guaranteed.

At the data level, the Yuanbao standalone app had 57.35 million monthly active users (QuestMobile, March 2026)—less than one-fifth of DouBao’s (note: Yuanbao’s total platform MAU, including the WeChat-embedded version, reached 114 million in February 2026, but daily active users dropped significantly after the Spring Festival red packet campaign). ByteDance CEO Liang Ruibo designated “Climbing New Peaks” as the annual keyword at the 2026 all-hands meeting, with the core goal of focusing on building the “DouBao/Dola” AI assistant app. Qwen is deeply integrated into Alibaba’s core scenarios, including Amap, Ant Pay, Fliggy, and Ele.me. The competitive dynamic has shifted from “which model scores higher” to “which agent can connect to a broader range of offline services and deliver a more complete execution loop.” Tencent’s response is not direct competition in the standalone app space, but rather at the ecosystem level.

WeChat's organic growth has entered a plateau, with 1.432 billion monthly active users and a year-over-year growth rate of just 2%, indicating that its domestic user base has reached its ceiling. Usage duration is more noteworthy: third-party estimates show WeChat averages about 85 minutes per day, having been surpassed by Douyin’s approximately 93 minutes. Daily posts on Moments have significantly declined from the 2021 peak, as user attention shifts from social interaction and messaging to short-form video and AI-native tools. This trend is reflected in financial data: Tencent’s Q1 social networking revenue amounted to RMB 31.9 billion, a 2% year-over-year decline.

A more direct financial signal is that Tencent’s AI capital expenditure in Q1 reached RMB 31.9 billion, a 16% year-over-year increase, while the quarterly net loss from new AI products reached RMB 8.8 billion, annualizing to approximately RMB 35 billion, or nearly RMB 100 million per day. Tencent’s management indicated that domestic AI chips will arrive progressively each month in the second half of the year, and capital expenditures will “increase significantly.” Spending growth far outpaces revenue growth, and Tencent must convert its massive AI investments into commercial returns. The WeChat AI Agent is the product closest to achieving this goal.

Why WeChat? WeChat spent eight years building several key components—today, it’s hard to find another company that can assemble all of them simultaneously.

Covering 108 niche industries and millions of mini-programs, standardized interfaces exist for nearly every aspect of daily life—transportation, dining, healthcare, government services, and more. Every merchant integrated into the WeChat ecosystem exposes programmable APIs. For AI agents, this means directly calling these interfaces, eliminating the need to guess button positions on screens as GUI-based solutions do. Precise, efficient, and with costs and error rates orders of magnitude lower.

In addition, within the user’s identity-based payment system, 1.4 billion users are verified under real-name registration, with their social networks deeply integrated with their payment accounts. From the very first second it receives a task, an AI Agent has full context: who the user is, their spending history, who can provide the service, and how to complete the payment. This is not an issue of model capability, but of ecosystem completeness. Doubao can help users find a coffee shop, but it cannot complete the payment—it lacks a payment license and cannot access the user’s identity information or bank card details.

Liu Chiping positioned the WeChat ecosystem as an "ideal assistant" during the earnings call—a system capable of understanding user needs and executing tasks within a closed loop. The WeChat AI Agent is not a newly created capability from scratch; rather, it connects three existing assets—mini programs, WeChat Pay, and the identity system—that have been in place for eight years using AI. By leveraging existing resources to drive new growth, WeChat is strengthening its current moat rather than digging a new one.

But whether the advantages can be realized depends on three variables.

Model. The Hunyuan 3.0 has a parameter scale of 295B, placing it at a medium-to-upper level within the industry, but a noticeable gap remains compared to the leading models. According to reports, the WeChat team has not fully bet on its self-developed Hunyuan model; it has already been testing models from Zhipu, Alibaba, and DeepSeek, and has also experimented with self-developed smaller models. The core challenge in adopting external models is not technical selection, but rather defining the authorization boundaries for WeChat’s internal data. With a user base of 1.4 billion, this issue has no simple solution.

Former OpenAI researcher Yao Shunyu is a key variable. He joined Tencent at the end of 2025 with significant authority to lead the upgrade of Hunyuan. It is reported that Hunyuan has indeed made considerable progress, but catching up still requires time. Zhang Xiaolong, head of WeChat, is well known in the industry for his high standards on product maturity: if features do not meet launch criteria, timelines may be adjusted at any time.

Computing power. Under chip export controls, Tencent was unable to stockpile sufficient quantities of NVIDIA’s high-end GPUs, and domestic chip production capacity remains in a tight balance. The computing power consumption of AI agents is on a completely different scale from ordinary conversations; a single natural language instruction from a user requires intent recognition, task decomposition, and multiple sequential micro-program calls, each step consuming vast numbers of tokens. Once scaled to 1.4 billion users, the inference cost would be enormous.

Liu Chiping stated at the Q1 earnings call that addressing inference-side computing power requirements requires a combination of strategies. Management also indicated that more domestic AI chips will be gradually deployed month by month in the second half of the year. However, more challenging than supply is the business model: offering services for free accelerates losses; charging for them contradicts WeChat’s longstanding policy of not monetizing core features. Doubao is expected to launch a paid subscription system by late June, and the industry is awaiting its commercial validation.

Developers. This is the most unassuming yet most structural variable. When AI agents directly assist users in invoking mini-programs to complete services, developers will lose users’ active visits and browsing. Much of the current mini-program business model relies on the user’s “browsing” journey: the homepage, product detail page, checkout page—each step enables conversion and ad monetization. With agents bypassing directly to the payment stage, these intermediate steps are drastically compressed. Redesigning developer incentive mechanisms is the core question determining whether the agent ecosystem can succeed.

In the January 2026 employee meeting, Ma Huateng said: “We will not control all entry points; we only provide the underlying connectivity. This approach is more scientific and reasonable, gives our ecosystem partners greater confidence and acceptance, and is therefore more sustainable.”

However, when it comes to reallocating benefits among millions of developers, developers' choices will directly shape the direction of the ecosystem. A more subtle variable is how companies with independent super apps—such as Alibaba, ByteDance, Pinduoduo, and JD—will position their investments within WeChat Mini Programs: will they tighten APIs and degrade user experience, or gradually migrate core services back to their standalone apps? Over the past year, scattered signs of such trends have already emerged. This is not a technical issue, but a redefinition of industrial ecosystem collaboration models.

How will the industry landscape be reshaped?

The main battlefield of AI competition has shifted from model benchmark scores to which agent can connect to more offline services, cover broader scenarios, and form a more complete execution loop. Liu Chiping specifically noted on the earnings call that daily active users are no longer the sole metric for measuring the commercial value of applications in the AI era. Tencent does not want to compete on others' rules—it aims to redefine the standards of measurement.

Meituan’s move provides a perspective on this. Wang Xing announced that Meituan’s AI assistant “Xiao Mei” will integrate with Tencent’s Yuanbao, allowing users to trigger local life services such as food delivery simply by stating their needs within Yuanbao. This marks the first experiment in AI-level interoperability between super apps. If WeChat’s AI agent adopts this model—integrating with Didi, JD.com, and Pinduoduo—it will no longer be merely Tencent’s agent.

But only if other companies are willing to open their doors. Qwen is deeply integrating Ant Pay, Amap, and Ele.me, while Doubao is connecting with e-commerce and local life scenarios. No major player will willingly cede the position of “user intent distribution” to WeChat. In the mini-program ecosystem, the experience of some leading companies’ mini-programs has already begun to lag behind their standalone apps—not necessarily due to neglect, but as a defensive strategy before the arrival of AI agents.

Tencent’s internal division of labor is already in place: Hunyuan handles the underlying engine, while CodeBuddy, WorkBuddy, QClaw, and OpenClaw experiment and accumulate experience in their respective scenarios, with the WeChat Agent ultimately serving as the unified endpoint. First, let scenario-specific agents pave the way, then consolidate everything under the WeChat Agent.

Is WeChat's AI Agent about catching up or upgrading? On the surface, it’s a product evolution—but what truly drives it is keeping the daily needs of 1.4 billion users within the WeChat ecosystem, from sending messages to handling tasks.

When a user opens Doubao and says, “Help me order a coffee,” the traffic entry point has already begun to shift. Even if the coffee shop ultimately uses WeChat Mini Program, the intent to “order coffee” no longer passes through WeChat. Control over the intent has quietly shifted within a single sentence. This is precisely why WeChat must develop AI Agents—to keep user service requests within WeChat as much as possible at the moment they arise, rather than allowing them to flow to Doubao, Qwen, or other external platforms. Even further, users no longer need to “go somewhere to find a service”—just speaking a sentence within WeChat is enough.

Fifteen years ago, WeChat redefined social interaction with voice messages. Nine years ago, it redefined app distribution with mini-programs. Both leveraged existing ecosystems to create entirely new experiences—and both succeeded.

This time, WeChat is attempting to redefine the connection between people and services using AI agents—the indispensable daily entry point for 1.4 billion users, built on millions of mini-programs with standardized interfaces. But in 2011 and 2017, WeChat was creating new growth; this time, it’s driving a structural upgrade within a saturated market, making the difficulty and complexity incomparably greater.

Reports indicate that users can simply swipe right on the main interface to open the AI chat window, transforming it from a social entry point into a service entry point. However, how successfully this step is executed isn’t just Tencent’s concern—every super app in China is watching to see whether WeChat’s AI integration can pave the way. (This article originally appeared on the Titanium Media APP; author: Jia Ywei)