Exclusive | Xu Can from Tencent Hunyuan has been transferred to WeChat WeLM, and WeChat AI has entered an accelerated development stage.

By | Li Zhaofeng

Edited by | Zhang Yuxin

Intelligent Emergence has learned from relevant sources that Xu Can, a senior researcher at Tencent Hunyuan, recently joined Tencent's WeLM team to engage in R&D work related to large language models.

Multiple insiders disclosed that before Yao Shuny took over the R&D of Tencent's large language model, Xu Can served as the person in charge of one post-training direction in the Hunyuan LLM system. It is worth noting that this job transfer took place at the stage when WeChat continued to strengthen its self-developed large model and Agent capabilities. In June, WeChat's native AI assistant "Xiaowei" has entered a small-scale gray test, which can operate WeChat's native functions through text or voice, and call mini-programs to complete service tasks.

Public recruitment information also shows that WeChat is currently recruiting researchers for WeLM reasoning optimization and Agent directions. Xu Can's joining may further supplement WeLM's R&D strength in training, reasoning and complex task execution.

This senior researcher with a background from Microsoft Research Asia has long been engaged in research on synthetic data, instruction data construction and large model post-training. His accumulation especially in reinforcement learning feedback, automatic evaluation and training data iteration is highly consistent with WeLM's current direction of strengthening model training and complex task capabilities.

Insiders said that Xu Can's experience in post-training work at Hunyuan in the past two years may be exactly what WeLM values.

01 Key Member of Hunyuan LLM, Focusing on Agent and Post-training Directions

This job change at Hunyuan is closely related to the technical route adjustment and personnel turnover of the team over the past year.

Xu Can's popularity in the large model industry can be traced back to the WizardLM project he created during his time at Microsoft. In April 2023, he published the WizardLM paper as the first author, proposing the Evol-Instruct method: let the large model automatically increase the difficulty, constraints and reasoning steps of existing tasks, and then use the generated complex instructions to train the model. This method was subsequently applied to projects such as WizardCoder and WizardMath, and gradually expanded to fields including code, mathematics, model evaluation and reinforcement learning.

In May 2025, Xu Can announced that he would leave Microsoft to join Tencent Hunyuan, and it is reported that he had already been on the job for half a year at that time. Before him, Hu Han, the main author of Swin Transformer (who recently left Tencent), also joined Hunyuan from Microsoft Research Asia in early 2025. Since then, the team represented by researchers with Microsoft backgrounds such as Hu Han and Xu Can began to use the signature "Hunyuan X", whose research covers multimodal understanding, visual reasoning and large language models, and was an important technical force within Hunyuan at that time.

Intelligent Emergence learned that during this period, Xu Can assumed important responsibilities in the direction of large language model post-training.

In the second half of 2025, with Yao Shuny joining Tencent and taking over related businesses, the original technical and organizational structure of the Hunyuan team began to change rapidly, and Xu Can's responsibilities were also adjusted.

In December of the same year, Tencent officially announced the restructuring of its large model R&D architecture. Yao Shuny concurrently served as the head of the AI Infra Department and the Large Language Model Department, and reported to Lu Shan and Martin Lau as the department head; Wang Di continued to serve as the deputy general manager of the Large Language Model Department and reported to Yao Shuny, and Xu Can also turned to the post-training direction at this time.

Intelligent Emergence learned that after Yao Shuny joined the company, Hunyuan continued to introduce researchers from teams of various large fundamental model manufacturers. In the post-training field alone, at least two core members from a large tech firm joined Hunyuan, and advanced different technical directions in parallel with Xu Can's team.

In the second half of 2025, some public reports described Xu Can as the main person in charge of Hunyuan post-training. But in fact, Hunyuan's post-training business had been split into multiple teams at that time, and Xu Can was in charge of one of the directions. According to Intelligent Emergence, several teams in the post-training team once explored directions such as data organization, system prompt words and product engineering paths.

After several adjustments to Hunyuan, Xu Can's public social media recently showed that his identity is Tencent Hunyuan Principal Researcher, that is, senior researcher, and he still often appears as an author in Hunyuan-related papers this year.

It is understood that WeLM has long retained independent R&D capabilities for language models and reinforcement learning, and also requires experience in synthetic data, reward models and post-training. The technical experience Xu Can has accumulated over the past few years can just find new application space in this team.

02 Beyond Hunyuan, WeLM's Independent Exploration on Large Models

This talent flow happens to fall at the time point when Tencent's two sets of large model R&D systems operate in parallel.

After Yao Shuny took over Hunyuan, Tencent began to further integrate the AI R&D resources scattered in various internal teams. In December 2025, Tencent Technology and Engineering Group (TEG) newly established the AI Infra (Artificial Intelligence Infrastructure) Department, the AI Data (Artificial Intelligence Data) Department and the Data Computing Platform Department, and Yao Shuny was in charge of the AI Infra Department and the Large Language Model Department at the same time.

A few months later, Tencent AI Lab, which had been established for nearly ten years, was dissolved. Some researchers were merged into the Large Language Model Department, and some others were transferred to the Industry-University-Research Cooperation Center. This round of adjustment further brought model R&D, data, training infrastructure and original basic research forces into the Hunyuan system led by Yao Shuny.

However, in Tencent's fundamental model R&D organization, Weixin Group (WXG) still retains a relatively independent R&D team.

In 2022, the WeChat team released the Chinese pre-trained language model WeLM. The paper disclosed that WeLM has 10 billion parameters, outperforms multiple models of the same scale on 18 Chinese tasks, and some of its performances can match models with 25 times more parameters, and it once opened model trial applications.

In August 2025, the WeChat team also publicly introduced WeChat-YATT, an RLHF training framework for large language models and multimodal models. This framework mainly solves problems such as controller scaling, dynamic sampling and GPU resource scheduling. The paper stated that it has been used to train models that support WeChat product functions and serve large-scale users.

At the same time, some large model technologies have been applied to the recommendation and advertising systems of Channels and Moments, but it is worth noting that not all the models used in them come from the WeLM system. For example, LEADRE deployed in the ad recall link of Channels and Moments uses the 1-billion-parameter Hunyuan model at the bottom.

By 2026, technical documents disclosed by Tencent show that the new generation of WeLM adopts a highly sparse MoE architecture. One of its basic versions has a total of about 80 billion parameters, and activates about 3 billion parameters for each inference; the team then formed a variant with about 130 billion total parameters and about 4.9 billion activated parameters through deep expansion.

On July 9, 2026, the WeChat AI team also published research on Hidden Decoding, disclosing two models WeLM-HD4-80B and WeLM-HD4-617B. This method adds internal calculation for each Token without expanding the original Transformer backbone, and applies this type of length dimension expansion method to MoE models with more than 100 billion parameters for the first time.

After Xiaowei entered the internal test, WeLM also began to get public attention. Many interpretations believe that WeLM is a "low-cost small model" specially trained for WeChat.

However, according to analysis by industry insiders, the WeLM team aims to drive the thorough transformation of WeChat with AI, including but not limited to scenarios such as chat, social interaction, and mini-program services.

This means that the WeLM team not only needs to develop low-activation parameter models to improve product deployment efficiency, but also explore large models of the 600-billion-parameter level to continuously break through the capability ceiling.

In this context, WeLM and Hunyuan are both collaborators and competitors at the resource level. The two teams have overlapping phased exploration work, and similar talent flows will also occur.

Behind the Two Fundamental Model Teams is the Competition for C-end AI Entrances

Yuanbao, Tencent's main C-end AI product, launched the Yuanbao Pai feature at the beginning of this year, which brought the Hunyuan and WeChat systems together at the product layer.

At that time, Yuanbao Pai tried to introduce AI into multi-person social scenarios, supporting functions such as group chat summary, image generation, and shared entertainment. The WeChat team also rarely released some communication resources, and Yuanbao Pai invitations could be spread through WeChat friends, group chats and Moments.

For Yuanbao, this is equivalent to obtaining Tencent's most core relationship chain entrance; for WeChat, this is also a product experiment to observe whether AI can reorganize social relationships.

However, under the background of WeChat's "limited" resource sharing, Yuanbao did not grow into the expected Tencent C-end breakthrough.

Looking only at the Spring Festival period, the gameplay of Yuanbao Pai and red envelopes once rushed to a traffic high. Tencent disclosed that as of February 18 this year, Yuanbao's daily active users exceeded 50 million, and its monthly active users reached 114 million.

However, this round of growth dividends did not continue to settle. QuestMobile data shows that after the red envelope event ended, Yuanbao's daily active users quickly fell back to the level before the Spring Festival; by March 2026, the monthly active users of the independent Yuanbao app was 57.35 million, failing to enter the camp of 100-million-level AI applications, and only three products including Doubao, Qianwen and DeepSeek had monthly active users exceeding 100 million.

At the same time, WeChat quickly launched the long-rumored native AI assistant "Xiaowei". It adopts a multi-model collaborative route: some industry internal test feedback shows that the main model of "Xiaowei" is WeLM developed by the WeChat team, and some answers will call DeepSeek.

In June this year, the WeChat Open Platform began to allow mini-programs to access the WeChat AI ecosystem: in the automatic mode, the platform can read the mini-program source code and analyze the page structure, allowing WeChat AI to operate directly; in the development mode, developers can also encapsulate business capabilities as skills that can be called by WeChat AI after being reviewed by the platform.

Judging from the functional form of "Xiaowei", Agent is the core capability for WeChat to connect AI to chat, social interaction and mini-program services, which is related to whether AI can truly understand requirements and complete operations. Xu Can's accumulated experience in post-training and Agent directions may help WeLM strengthen its capabilities in task planning, tool invocation and complex task execution.

WeLM is positioned to focus on balancing user privacy, WeChat scenario adaptation and cost efficiency. Facing WeChat's huge and complex user scenarios, the future WeChat Agent not only needs to understand user intentions, but also plan tasks, invoke tools and complete operations stably between different services. This will put forward higher requirements for the model's post-training, reward design, task evaluation and execution stability.

In this context, it seems reasonable that the WeLM team values candidates' experience in the field of general large model post-training.

In fact, behind Hunyuan and WeLM, it is not difficult to find that Tencent has two sets of organizational logics for AI entrances: Hunyuan has more centralized computing power, data and model R&D resources; WeLM has WeChat's native entrance, content ecosystem, mini-program services and real user feedback.

However, both of these two logics require a strong technical base to support.

Intelligent Emergence learned that previously, the WeLM and Hunyuan teams have shared part of the computing power resource allocation. This transfer of personnel to WeLM further indicates that the synchronous operation of these two large model R&D systems will continue at the model and product levels.