On June 8, 2026, the WeChat Developer Platform announced that WeChat AI had entered the internal testing phase. This AI assistant, integrated into the WeChat ecosystem, allows users to directly invoke, access, and operate mini-programs through natural language dialogue. The open platform provides two access modes: automatic mode allows authorized platforms to read the mini-program's source code, enabling AI to directly operate the page without additional development; development mode allows developers to build skills independently, which are then reviewed and made available for AI to use. The terms of service also state that "WeChat AI" may be a temporary name, the final name has not yet been determined, and access is optional and will not affect the normal operation of existing mini-programs.

This marks the first time WeChat has opened its mini-program ecosystem to AI at the dialogue entry point. The background is that Tencent's self-developed Hunyuan large-scale model has already ranked among the top tier in domestic public benchmark tests, and the Yuanbao App's monthly active users exceeded 100 million after the surge in popularity of its 2026 Spring Festival red envelope campaign. WeChat AI's internal testing represents the latest step for Tencent AI from technological reserves and independent product verification to the delivery of a super application. The automatic mode requires developers to submit their source code; how many developers will this low-barrier-to-entry path attract, and what ecosystem conflicts of interest will it encounter? These are questions that the internal testing phase needs to answer.

Opening the dialogue layer in the mini-program ecosystem

WeChat AI's two access modes are aimed at completely different developer groups.

The design logic of the automatic mode is straightforward: the authorization platform reads the mini-program's source code during the submission process, automatically analyzes the page structure, and allows AI to directly manipulate the page without additional development. A small game team of only two or three people doesn't need to employ AI engineers or understand the Agent protocol; they simply need to select authorization, and their ordering mini-program and utility applications can be invoked by WeChat AI.

According to data disclosed by WeChat Open Class in January 2026, the WeChat Mini Game ecosystem has gathered over 400,000 developers, 80% of whom are small teams of fewer than 30 people. In 2025, the total daily active users exceeded 100 million, and monthly active users exceeded 500 million. This scale of supply is a unique moat for WeChat AI. ByteDance's Doubao or Alibaba's Tongyi Qianwen can create a standalone app and open APIs, but they don't have a mini-program ecosystem with over 100 million daily active users that can be directly integrated. WeChat AI's automatic mode essentially exchanges technological convenience for large-scale integration, allowing the vast majority of the 400,000 developers to get on board at zero cost.

The development mode provides customization options for service providers with complex business logic. Developers can independently build skills based on their own business characteristics, which can then be evaluated and approved by the platform for use by WeChat AI. Both modes can be used simultaneously and are not mutually exclusive.

The wording "name undetermined" and "optional behavior" indicates that the WeChat team still has reservations about the product's positioning. The main task during the internal testing phase is to test the technical pipeline and observe developer feedback. However, the automatic mode has touched a sensitive point: source code authorization. Some developers have expressed concerns on the WeChat Open Community, with core issues focusing on several aspects—how to ensure code asset security after the platform reads the source code, whether AI directly manipulating the page will cause existing tracking and ad display logic to fail, and how to assign responsibility if AI misoperation leads to user losses. There are currently no publicly available details regarding these issues.

After achieving second place in basic skills domestically, Hunyuan chose to delve deeper into the field.

WeChat AI needs more than just a chatbot; it needs an agent-based platform that can understand page structure and accurately execute commands. This platform is Tencent's Hunyuan Big Model.

In March 2025, SuperCLUE, a benchmark for evaluating large-scale Chinese learning models, released a report in which Tencent's Hunyuan flagship version ranked second in China in the basic model ranking, second only to ByteDance's Doubao. However, it ranked first in China in the application capability dimension, leading in sub-categories such as text understanding and creation, instruction compliance, and agent capabilities. ScienceNet, in its report, pointed out that Hunyuan's performance in the "practicality" dimension was better than its basic capability ranking. At the same time, Hunyuan Turbo S entered the global Top 15 of the international evaluation Chatbot Arena for the first time.

Hunyuan maintains a quarterly version iteration pace. In April 2025, hunyuan-turbo was updated, and in July, the flagship version TurboS was launched, enhancing its reasoning capabilities. In April 2026, the Hy3 preview version was released, with the official claim of a 40% improvement in reasoning efficiency. According to Tencent Cloud product documentation, older versions such as HY 2.0 are scheduled to be discontinued starting June 26, 2026.

This pace is much slower than that of ByteDance and Alibaba. ByteDance's Doubao and Alibaba's Tongyi Qianwen have maintained a near-weekly model release frequency over the past year, while Hunyuan has consistently released a major version update every quarter. Tencent's management has previously made public statements about "slow and steady wins the race," explaining from a technical perspective that the requirements for stability and low latency in the Agent era are far higher than in the Dialogue era, and frequent switching of underlying models would prevent developers from doing engineering adaptations. The scenarios that WeChat AI needs to invoke include operations involving funds and sensitive information such as ordering, payment, and appointments, where the determinism of the model output is far more important than creativity.

Regarding resource allocation, Tencent President Martin Lau disclosed at the 2025 annual report communication meeting that the company would invest 18 billion yuan in the R&D of new AI products in 2025, and that this investment would at least double in 2026. According to the meeting content relayed by The Paper, Lau also stated that the next core plan is to create a dedicated AI agent within WeChat, connecting mini-programs, social networking, and payment across the entire ecosystem. The doubling of investment without accelerating the release schedule indicates that funds are flowing more towards infrastructure reconstruction and data quality improvement, rather than rushing to release products.

Hunyuan's leading application capabilities resonate with the scenario requirements of WeChat AI. A model with a higher base model ranking but weaker agent capabilities may actually be less effective than Hunyuan in WeChat AI scenarios. Tencent has chosen a path that avoids parameter competition and focuses on practical dimensions, a path whose logical consistency began to show during the WeChat AI beta testing.

Daily active users exceeded 50 million during the Spring Festival, and then what?

Prior to the WeChat AI beta test, the C-end verification task for Tencent AI was undertaken by the Yuanbao App.

Yuanbao's growth curve exhibits a clear pulse characteristic. According to data from QuestMobile, cited by CNR.cn, Yuanbao ranked 12th in monthly active users (MAU) in January 2025, and rose to 3rd by December 2025, second only to Doubao (MAU 226 million) and DeepSeek (MAU 135 million), with a compound annual growth rate of 27.8%.

During the 2026 Spring Festival, Yuanbao experienced explosive growth. Official data released by Tencent shows that Yuanbao's peak DAU exceeded 50 million, reaching 40.54 million on New Year's Eve, with MAU reaching 114 million. The Shanghai Securities News reported that this growth was mainly driven by user acquisition through the social network of the red envelope campaign.

However, the data dropped rapidly after the Spring Festival. QuestMobile monitoring shows that in April 2026, Yuanbao's normal DAU was about 9 million, while Doubao's DAU was about 140 million and Qianwen's was about 30 million. The peak-to-trough difference was nearly 5 times, showing obvious pulse-like growth characteristics. No publicly available data is available on the DAU to MAU ratio, making it impossible to make a definitive judgment on user stickiness.

Yuanbao's role in Tencent's AI strategy is that of a "consumer-side validation for independent products." It proves two things: first, Tencent has the ability to leverage WeChat's social network to reach hundreds of millions of users; second, users attracted by red envelopes are not retained. At the earnings conference, Liu Chiping stated that Yuanbao's Spring Festival promotion exceeded expectations, and the next step is to focus on optimizing core capabilities such as voice dialogue. This statement itself demonstrates that the team understands that user retention is the core issue for the next stage.

The experience of Yuanbao's rapid growth explains why WeChat AI chose to integrate natively within the super app, rather than continuing to push a standalone app. Standalone apps require users to actively open them, and retention relies on push notifications and activities; native integration, on the other hand, binds users to specific scenarios. When users need to order food, pay bills, or check their packages, WeChat AI is already in the conversation flow. These are two completely different retention logics.

Every mini-program can be "lobster-ified," but service providers are afraid of being short-circuited.

The product direction of WeChat AI was already clearly outlined in Ma Huateng's public statement in March 2026.

During the 2025 annual report communication meeting, Ma Huateng (Pony Ma) first discussed the concept of "lobster farming." The "lobster-like" applications he referred to are AI agents with a "human-like" quality, capable of autonomously performing tasks rather than simply answering questions. Ma Huateng stated that these applications have inspired the planning of WeChat AI: in the future, every mini-program can undergo intelligent, "lobster-like" transformation.

The core of this analogy is to shift AI from a conversational tool to a task executor. If WeChat AI were merely a chatbot, it wouldn't need to read source code or manipulate pages. The existence of the automatic mode indicates its role is to complete cross-mini-program tasks for users: ordering a coffee, paying utility bills, booking a hospital appointment, or launching a mini-game. Users don't need to know which mini-program provides these services; they only need to say a sentence to WeChat AI.

However, Ma Huateng proactively addressed the conflict of interests within the ecosystem at the same meeting. He pointed out that ecosystem service providers are worried about being "short-circuited" and "channelized" by AI agents. If a user tells WeChat AI, "Order me a latte," and the AI ​​directly calls an atomic service of a coffee mini-program to complete the transaction without the user ever accessing the merchant's page, then the merchant's advertising space, brand exposure, and user base will all be wiped out. Service providers will not accept this outcome.

This is the core contradiction in WeChat's AI product design. The more efficient centralized scheduling is, the weaker the decentralized traffic sovereignty of merchants becomes. The two access modes themselves do not resolve this contradiction; they are merely entry point designs. The true balancing mechanism, such as traffic allocation rules, the relationship between atomic services and merchant pages, and the data visibility of service provider backends, has not yet been publicly disclosed. Ma Huateng's original words were "we must balance centralized scheduling with decentralized traffic protection," but how exactly to balance them has not yet been answered during the internal testing phase.

The three lines are in place, but the third step has just begun.

With Hunyuan, Yuanbao, and WeChat AI progressing in parallel, Tencent AI's gradual path is logically self-consistent.

The underlying layer doesn't focus on the fastest model, but rather on the most stable agent foundation. Hunyuan's top ranking in SuperCLUE application capabilities in China supports WeChat AI's need for precise operations. The middle layer uses a standalone app to seamlessly integrate social networks for user acquisition and basic user experience; Yuanbao's MAU exceeding 100 million during the Spring Festival validates the leverage effect of WeChat's traffic pool on AI products. The upper layer features native integration within the super app, using scenarios to reduce retention pressure. WeChat AI's internal testing directly faces 400,000 developers and the mini-program ecosystem with over 100 million daily active users.

However, whether the perception among end-users has shifted can only be assessed as "partially complete" at present. Yuanbao's hundreds of millions of monthly active users (MAU) mainly come from red envelope promotions, with a typical DAU of around 9 million, significantly lower than Doubao's 140 million. WeChat AI is still in internal testing, and ordinary users cannot yet perceive its impact. There is still a significant gap between Tencent AI's public perception and its technological level.

Whether WeChat AI can bridge this gap depends on three variables. First, whether the source code trust issue of the automatic mode can be resolved on the developer's side, which determines the scale of supply-side integration. Second, whether the centralized and decentralized traffic allocation rules are acceptable to service providers, which determines whether a balance can be achieved in the ecosystem's interests. Third, whether the accuracy of AI operations and the attribution of responsibility can reassure users when placing orders, which determines the depth of C-end usage.

The three lines being in place are a prerequisite, but whether they can form a chain of "Hun Yuan ensuring reliability, Yuanbao verifying user habits, and WeChat AI delivering the final experience" still needs at least two quarters of publicly available data to verify. As Ma Huateng said at the earnings conference, "AI is a marathon, not a sprint," and the WeChat AI internal testing is just a marker point in this marathon; there's still a long way to go before the finish line.