Post Metaverse Training, Backbone Xu Can Transfers to WeChat WeLM: WeChat AI Enhanced Agent

According to Dongcha Beating monitoring, Tencent Mixologist senior researcher Xu Can has recently transferred to the WeChat WeLM team to work on large-scale model development. Prior to Yao Shunyu taking over Mixologist's large language model development, Xu Can was responsible for a direction of Mixologist's LLM post-training.

Xu Can is the creator of Microsoft's WizardLM project and the first author of the paper. He proposed Evol-Instruct, which allows large models to automatically generate more complex instruction data for training. He joined Tencent Mixologist in 2025 and has long been researching synthetic data, instruction training, and large model post-training.

This job transfer coincides with WeChat's reinforcement of WeLM and Agent. A public job posting shows that the WeLM team is recruiting researchers for inference optimization and Agent, including researching "training strategies for a general Long-Horizon LLM Agent." WeChat had previously released the RLHF training framework WeChat-YATT for training models that support WeChat products.

WeChat's own model is also rapidly expanding. The July release of the Hidden Decoding research has achieved WeLM-HD4-80B and WeLM-HD4-617B. Meanwhile, WeChat's native AI assistant "Xiao Wei" is undergoing small-scale grayscale testing, able to operate WeChat's native functions and call mini-programs. Xu Can's post-training experience aligns well with WeChat's current direction of enhancing Agent's capabilities for complex tasks.