WeChat Deploys 80B WeLM for Xiao Wei While 617B Model Remains in Development

WeChat integrates the 80B-parameter WeLM into its Xiao Wei AI agent for search and mini-program tasks. The larger 617B MoE model is still under development for complex agent functions.

Woofun AI reports that the WeChat team has revealed two configurations of its self-developed WeLM large model. The WeLM-80B variant, featuring 80 billion total parameters with 3 billion active at any time, is now integrated into the native AI agent Xiao Wei to manage chat search, invoke WeChat functions, and serve mini-programs.

The larger WeLM-617B model, utilizing a Mixture of Experts architecture with 617 billion total parameters and 23 billion active per instance, remains under development. Tencent indicated in its Q2 earnings report that Xiao Wei is undergoing limited Grayscale testing within WeChat, leveraging WeLM for privacy, scenario specificity, and inference efficiency. These models were previously detailed in the WeChat team's July Hidden Decoding paper.