WeChat's Xiaowei Agent Runs on WeLM — a Sparse MoE Model Quietly Growing to 617 Billion Parameters

WeChat has started gray-testing Xiaowei, a native AI assistant embedded inside the app, powered by WeLM, the WeChat AI team's long-developed large language model. WeLM has moved to sparse MoE architecture, with a 617-billion-parameter version activating just 23 billion per token, and the team is exploring Hidden Decoding as a new scaling direction.