WeChat's Xiaowei agent runs on a secret 617B MoE - Tencent's quiet frontier bet

WeLM, the model powering WeChat's AI agent, has quietly scaled to 617 billion parameters with an undisclosed decoding mechanism. No benchmarks, no papers - just a billion-user deployment.

In plain terms: Tencent's Xiaowei agent, being gray-tested inside WeChat, runs on WeLM - a sparse mixture-of-experts model that has quietly grown to 617 billion parameters. No public benchmarks. No papers. Just a large-scale gray-test on one of the world's highest-traffic messaging platforms.

The fact

According to WeChat AI team disclosures reported by Pandaily, WeLM has reached 617 billion parameters in its sparse MoE version, activating 23 billion parameters per token. Tencent has never published WeLM. The model powers Xiaowei, WeChat's integrated AI assistant - currently in gray-testing across WeChat's 1.3B+ MAU base.

Our read

The leaderboard-first narrative of the LLM race misses models like WeLM: frontier-scale, production-path, fully closed, and backed by a distribution moat that no API-first lab can match. Alibaba competes with Qwen via benchmarks and open weights. Tencent competes via embedding.

Under the hood: Sparse MoE architectures activate only a subset of parameters per token - WeLM activates 23B of its 617B per inference step. This is comparable to a ~23B dense model in compute terms, which explains how a model at this nominal scale can run in a consumer-facing chat interface at latency acceptable for messaging.

The "hidden decoding mechanism" disclosed in the report likely refers to speculative decoding or an early-exit routing strategy optimized for latency - the kind of inference engineering that doesn't make papers but makes products fast.

Watch

Any public WeLM benchmark disclosure; whether Tencent open-weights WeLM (unlikely); how Xiaowei performs versus ChatGPT and Kimi as gray-testing expands.