52.78 billion yuan.

That's how much Tencent spent on compute capex last quarter - up a staggering 176% year-over-year, pushing free cash flow negative for the first time since its IPO. Management's explanation on the earnings call was blunt: the money all went to buying GPUs and hoarding compute.

Around the same time, a far more telling personnel move quietly landed: Xu Can, a senior Hunyuan researcher and first author of WizardLM who proposed Evol-Instruct, transferred to the WeChat group to join the WeLM team, leading post-training and agent development.

Put these two together and Tencent's AI roadmap comes into focus: Hunyuan is the regular army, WeLM is the special forces, and the real endgame hides inside WeChat, a super app with 1.6 billion users.

From WizardLM to WeLM: A Key Figure Moves

Who is Xu Can? In the LLM world, this name is bound to two technical labels: WizardLM and Evol-Instruct.

In April 2023, Xu Can published the WizardLM paper as first author, proposing Evol-Instruct - a method that lets models automatically add difficulty, constraints, and reasoning steps to existing tasks, then trains on the generated complex instructions. The method was later adopted by WizardCoder, WizardMath, and others, becoming a classic in synthetic data and post-training.

A Microsoft Research Asia background, deep experience in synthetic data, RL feedback, and automatic evaluation - all of it lands exactly on what WeLM needs most right now.

WeChat's WeLM is actually an "old project": it released a 10B-parameter Chinese pre-trained model back in October 2022, a full month before ChatGPT. For three years it vanished from public view while iterating in the shadows - between January and July this year, the WeChat AI team published four technical blogs, fully revealing WeLM's post-training, MoE architecture, and long-sequence approaches.

Xu Can's arrival fills the post-training and agent piece of the puzzle. Public job listings show WeChat is actively hiring WeLM inference-optimization and agent researchers.

Why WeChat? Agentifying the Super App Is the Endgame

To grasp the weight of this transfer, you first need to see Tencent's AI predicament.

In the chatbot race, Tencent's Yuanbao trails Doubao by an order of magnitude; on base models, Hunyuan is still chasing SOTA. Grinding further down those two tracks doesn't offer great odds.

But Tencent holds a card no one else has: WeChat. 1.6 billion users, hours of daily attention, a complete ecosystem spanning payments, social, content, and services - if WeChat itself becomes an agent entry point, that's the real killer app.

This June, WeChat's native AI assistant "Xiaowei" entered grey testing, letting users operate native WeChat features by text or voice and invoke mini-programs to complete service tasks. And Xiaowei doesn't run on Hunyuan - it runs on WeLM.

What does that mean? It means WeChat is walking a path parallel to - and even more aggressive than - the group's Hunyuan: not chasing the strongest base model, but chasing the most useful one inside the super app.

The financials show the same strategic tilt: last quarter Tencent's compute-related cash outflow exceeded 110 billion yuan, and management explicitly said the compute would go to "WeChat AI initiatives." Where the money flows, the strategy sits.

Xu Can's transfer isn't an ordinary personnel change. It's a signal: Tencent is betting its AI endgame on WeChat's agentification. Hunyuan keeps building the base, WeLM attacks the scenario - two legs walking, but the real growth engine is the latter.

When a 1.6-billion-user super app starts turning into an agent that can understand and get things done, Tencent's ticket to the AI era may be worth far more than anyone predicted.

See you tomorrow.