Tencent's trump card has never been large models.

The outside world may have always misjudged Tencent's AI: it's not that it's slow, but that it dares not be fast.

An exclusive scoop from a certain newspaper yesterday completely disrupted the landscape of the domestic AI race. Tencent is secretly testing a prototype of an AI Agent embedded in WeChat, planning to initiate the compliance approval process and then conduct a small - scale gray - scale test. As soon as the news came out, Tencent's Hong Kong - listed shares soared by 10%.

In the past two years, industry public opinion has unanimously judged that Tencent's AI has fallen behind. ByteDance, Alibaba, and Baidu have been running at full speed, competing for entry points, piling up computing power, and expanding application scenarios. However, Tencent has still been labeled as "conservative, lagging, and slow". But in the game among top Internet giants, it's never about who runs faster, but about who makes more accurate bets and holds on more steadily. Finally, Tencent is going to play its trump card of WeChat AI.

I. Tencent's AI Dilemma

The interaction logic of the WeChat AI Agent revealed by the media is extremely simple: swipe right on the WeChat home screen to evoke the intelligent dialogue panel. Users can complete the full - link life service fulfillment with just one natural language instruction, including finding stores, comparing prices, placing orders, and making payments, without manually opening any mini - programs or repeatedly jumping between pages.

This is fundamentally different from Yuanbao AI. Yuanbao is just an auxiliary tool, mainly for dialogue Q&A and search enhancement. It doesn't involve transactions, doesn't interfere with user behavior, and doesn't change the ecological rules. It's a typical "icing on the cake".

The WeChat AI Agent, on the other hand, is a complete underlying innovation. Its core ability is to schedule the entire WeChat and the mini - program business empire behind it on behalf of users.

For more than a decade since its establishment, WeChat's core positioning has always been a "super connector", a traffic container for nearly ten million mini - programs, and a platform for distributing traffic and building links. Users are the main body of all operations, actively searching, clicking, and fulfilling tasks. WeChat only acts as a rule - maker and traffic distributor.

After the implementation of the AI Agent, the situation is completely reversed. WeChat has been upgraded from a "traffic container" to an "intention operating system" and from "passive connection" to "active agency". It starts to take over users' decision - making and operations, monopolizing the full - link closed - loop from "intention to execution".

Ma Huateng once admitted that Tencent's AI layout was nearly a year behind. But the most misunderstood point in the industry is that Tencent's "slowness" is not due to lack of ability, but because it dares not and cannot be fast.

The AI products of ByteDance, Alibaba, and Baidu can quickly test, iterate, and even fail, because their AI entry points are independent, replaceable, and have a large tolerance for errors. However, WeChat can't.

Today's WeChat has gone beyond the scope of a social app and has become a real national digital infrastructure. With 1.4 billion monthly active users and a transaction scale of trillions of yuan per year, it covers all core life scenarios of Chinese people, including communication, payment, government affairs, travel, medical care, and utility bill payments. Any function embedded in WeChat will directly define the Internet usage habits of 1.4 billion users.

This is a double - edged sword. If it succeeds, WeChat will firmly lock in its position as the super entry point in the AI era, and its barriers will remain solid for another decade. But if it makes a mistake, the cost will be devastating.

AI hallucinations misjudging instructions, incorrect bill payments, privacy leaks, wrong service port adjustments... Any small loophole will escalate from a product experience problem to an Internet public opinion event and a public trust crisis. For WeChat, the premise of innovation is absolute safety and stability, which is a heavy burden that no other competitor needs to bear.

Under this zero - tolerance pressure, Tencent has remained restrained in the past two years: the Hunyuan large - scale model has been iteratively updated in a low - key manner without external hype; Yuanbao AI is limited to lightweight auxiliary functions; all the intelligent agents in its Lobster series are implemented in controllable scenarios such as enterprise services and developer tools, resolutely not touching the main WeChat ecosystem and not interfering with the core experience of C - end users.

This restraint was only broken when the user habits of competitors were gradually taking shape and ready to be harvested.

II. Encirclement by Competitors

The current AI strategies of the three major tech companies, although seemingly in different tracks, actually precisely target Tencent's weak points and form an encirclement.

ByteDance's Doubao takes the route of an AI - native traffic entry point. Relying on ByteDance's powerful recommendation algorithm and the mentality of young users, it quickly seizes users' fragmented time in the form of an independent app. With lightweight conversations and low - threshold experiences, it crazily cultivates users' new habit of "getting things done with one sentence". It is lightweight, has few burdens, and can iterate quickly. Its only goal is to seize the first interaction entry point of users.

Alibaba's Tongyi Qianwen focuses on the closed - loop of business transactions. It binds core scenarios such as e - commerce, travel, local life, and cultural tourism, and deeply integrates AI capabilities into the entire process of transaction fulfillment. It emphasizes practicality and implementation, firmly holding on to high - value business scenarios and strengthening its own transaction AI barriers.

The core advantage of these two lies not in technologically crushing Tencent, but in defining the user interaction rules in the AI era first.

In the past two years, countless users have developed a new perception: when they have a need, they don't need to operate manually, just talk to the AI to get things done. Conversation is operation, and intention is result.

This is a dimensionality - reduction blow to WeChat. If users gradually get used to completing service fulfillment on Doubao and Qianwen, and then look back at WeChat's cumbersome traditional operations of "search - click - jump - place an order", the irreplaceability of WeChat will instantly collapse.

WeChat won't be uninstalled, but it will be downgraded.

This is Tencent's deepest fear: not that the current user scale of competitors exceeds its own, but that opponents have preemptively captured the users' minds in the AI era. Once "conversational fulfillment" becomes an industry consensus, when WeChat enters the market, it will no longer be the rule - maker and can only follow passively. The status of the super entry point will completely become a simple social pipeline.

It is this external encirclement that has upgraded the WeChat AI Agent from a "long - term strategic plan" to the highest - level strategy of the Tencent Group, breaking the product restraint bottom - line that WeChat has adhered to for many years.

III. A Difficult Trump Card

The outside world only sees Tencent's boldness in playing its trump card, but fails to see that behind this trump card are multiple deadlocks in technology, computing power, ecosystem, and compliance. Just because the prototype can work doesn't mean it can be used well by 1.4 billion people. The implementation of the WeChat AI Agent is arguably the most difficult product engineering problem in the Internet.

First, computing power is a bottleneck, and the cost is astronomical.

The AI Agent is completely different from ordinary conversational AI. A single service requires multi - step reasoning, cross - ecosystem scheduling, and real - time data interaction, and the computing power consumption increases exponentially. For a national - level product with 1.4 billion monthly active users, the computing power demand after full - scale opening is a real "computing power black hole".

Due to overseas chip restrictions, Tencent had insufficient reserves of high - end GPUs in the early stage, and the substitution of domestic computing power is still in the adaptation and improvement stage. Liu Chiping also recently publicly stated that last year, due to hardware restrictions, the AI capital expenditure did not meet expectations, and this year, the AI investment will be doubled.

This means that Tencent is not unable to develop Agent technology, but it can't balance the cost of large - scale use by 1.4 billion users. The current gray - scale test is essentially waiting for the three curves of computing power supply, hardware adaptation, and cost control to reach a balance.

Second, adapting to the 7 - million - strong mini - program ecosystem is a hell - level project.

WeChat's vast mini - program ecosystem is Tencent's core barrier, but also the biggest drag on the implementation of the AI Agent. The development standards of millions of mini - programs are fragmented, the interfaces are not unified, the business logic is chaotic, and the permission systems vary.

For the AI Agent to achieve accurate fulfillment, it needs to complete the full - link operation of "natural language recognition - precise matching of mini - programs - adaptation to corresponding interfaces - completion of authorized payment". Any mismatch in any link will lead to service failure.

What's more troublesome is the conflict of ecological interests. The core income of most mini - program developers comes from page exposure, user clicks, and traffic conversion. Once the AI Agent automatically completes price comparison, order placement, and fulfillment in the background, directly skipping all display links, it is equivalent to depriving developers of traffic and income.

Tencent needs to ensure the efficient experience of the AI Agent and at the same time appease the interests of millions of ecological partners to avoid ecological turmoil. This balance of interests is a more difficult product management problem than technological R & D, which is also the core reason for Tencent to try the hybrid model of centralization and decentralization.

Third, compliance and privacy are red lines that cannot be crossed with zero tolerance.

The core of intelligent proxy fulfillment is to deeply call user data: chat records, geographical locations, consumption habits, payment information, and livelihood data all need real - time authorization, reading, and analysis.

With a user base of 1.4 billion, any problem of data abuse, privacy leakage, or permission overstepping will cross the regulatory red line. This is also the core reason why Tencent must initiate a strict compliance process and adhere to a small - scale gray - scale test - there is no room for error in this iteration.

Finally, let me say a few more words. In fact, the competition among big tech companies in AI is not about model parameters, algorithm capabilities, or conversational intelligence. The ultimate value of any AI intelligent agent is not chatting and answering questions, but implementation: placing orders, making payments, handling affairs, consuming, and providing services.

From this perspective, the launch of the WeChat AI Agent is more like a defensive battle: Tencent doesn't need its AI to be more intelligent than all competitors. It only needs to ensure that all users' life needs and service intentions are finally closed - looped within the WeChat ecosystem; preventing the outflow of users' intentions, traffic, transactions, and data is the core of winning the AI war.

But the real test actually starts right now: Does Tencent dare and can it really hand over the control of the trillion - level ecosystem to an uncertain AI black box?