WeChat Introduces WeLM Dual Model: 80B Model Empowering Mini Programs, 617B Version Still in Development

According to Perceiving the Future monitoring, the WeChat team has announced the two configurations of their self-developed WeLM large model. The WeLM-80B has a total of 800 billion parameters, with only 30 billion parameters activated at a time. It has now been deployed to the WeChat native AI Agent "XiaoWei," handling tasks such as chat search, invoking WeChat native functions, and Mini Program services.

The larger WeLM-617B is still under development, with a total of 617 billion parameters, activating 230 billion parameters at a time, using the MoE architecture. WeChat has stated that it will enhance general understanding and reasoning capabilities, mainly targeting more complex WeChat ecosystem tasks, including developing smart Mini Programs and generating tools for XiaoWei.

Today, Tencent also mentioned in its Q2 financial report that XiaoWei is undergoing small-scale gray-scale testing within WeChat, driven by WeLM customized for user privacy, WeChat scenarios, and reasoning efficiency.

The WeLM-80B and 617B had previously appeared in a July Hidden Decoding paper by the WeChat team. This further clarifies the product positioning of the two models: the 80B is already in practical service for XiaoWei, while the 617B is aimed at more complex Agent and tool invocation tasks.