Tencent Joins China’s AI Agent Race With ‘Top-Secret’ WeChat Project

TENCENT is secretly building a new AI agent for its hugely popular WeChat messaging app, in hopes of leapfrogging rivals like Alibaba Group and ByteDance in the race to dominate China’s domestic AI market, according to four people familiar with the project. The project could dramatically widen how AI agents, which take actions on behalf of people, are used in China, where WeChat is ubiquitous in people’s daily lives.

The new agent would connect with the millions of miniprograms—lightweight apps—running inside WeChat that provide scores of services, from booking taxis to ordering groceries, according to three of the four people. If it works, the agent could undertake those tasks on behalf of WeChat’s 1.4 billion monthly active users.

The company is treating the project, which has been in the making since at least the first half of last year, as a high-priority, top-secret initiative, according to the three people. The agent, which would be for WeChat’s domestic version known as Weixin in China, will appear as a chat in the list of conversations users can call up and interact with, two of them said.

Tencent’s move is emblematic of the stakes AI agents hold for tech titans. From Silicon Valley to China, companies are racing to launch AI assistants that can autonomously complete tasks for both workers and consumers, from coding to commerce.

The stakes are particularly high for Tencent, which has moved slowly to embrace AI. In contrast, Alibaba and ByteDance in the past months have taken steps to transform their chatbots into all-in-one agents that can perform daily tasks for consumers. Alibaba has connected its Qwen app with its e-commerce, online travel and mapping features and with affiliate Ant Group’s payment platforms, so the app can perform tasks like online grocery shopping and booking flights. ByteDance has also turned its Doubao app into an agent that handles e-commerce and other tasks. In addition, both Alibaba and ByteDance last month released new-generation AI models that they say are more suitable for handling complex, multistep tasks.

Tencent, however, can’t afford to disrupt the experiences of WeChat’s vast user base with half-baked technologies or features, according to two people familiar with Tencent executives’ thinking.

When Tencent decided to launch a new chatbot app in May 2024, it did so outside WeChat. The new app, Yuanbao, has not taken off. As of February, it had 109 million monthly active users, dwarfed by ByteDance’s Doubao, which had 315 million, and Alibaba’s Qwen, which had 202 million, according to Aicpb.com, a Chinese website that tracks AI products. Tencent last year added a Yuanbao chatbot feature to WeChat, but it serves more as a conversational AI assistant than an agent.

WeChat’s conservatism is a sharp contrast from the app’s early years. Some of its innovative features, such as hold-to-talk voice messages, launched in 2011, have inspired WhatsApp, which added the same function in 2013. When WeChat pioneered miniprograms in 2017, the fast, frictionless user experience quickly made the app a formidable alternative to the traditional App Store model, threatening Apple, and eventually turned it into a do-everything superapp that spans tasks way beyond messaging. WeChat’s miniprograms have been so successful that they have sparked copycat efforts from Microsoft, Snapchat and Google.

Hired From OpenAI

Tencent started ramping up its efforts last September, when it hired researcher Yao Shunyu from OpenAI. Tencent’s leaders empowered Yao to lead the development of the company’s AI models—known as Hunyuan or HY—and gave him the budget and latitude to poach talents from rivals such as ByteDance, The Information reported in November.

Even though the team that works on HY models is in the process of a makeover under Yao, the WeChat team, so far, has not been using HY models for the AI agent project, because they are still not among the top performers in the industry, three people familiar with the project said. The team has tested AI models developed by other Chinese companies, including Zhipu, Alibaba and DeepSeek, to power the WeChat agent, in addition to a small-size model developed by WeChat itself, two of the three people said. But working with other companies’ models means it will take longer to integrate and verify data stored on WeChat.

The WeChat team, led by the app’s legendary creator, Allen Zhang, has been working on its own AI model research and development. In January, WeChat AI published two papers on its official blog, the first about how to build more capable models with limited resources, and the second about post-training. Harvey Zhou, WeChat’s technology chief, who reports to Zhang, oversees the AI team, according to a person with knowledge of the matter.

Tencent plans to launch the feature for a trial among select users by midyear—a process known as gray-box testing in software development—and roll out to all users by the third quarter, according to a person involved in the project development. That timeline could still shift as WeChat won’t release the agentic features until they are solid, the person said.