The Painful Transition Period of Tencent Hunyuan: When the New Leader Takes Over, the Former Old Guard Steps Down
In the third year of the rapid, breakneck advancement of large models, the internal organizational structure of major tech giants has finally ushered in its first wave of substantive, in-depth restructuring.
Recently, news about the organizational structure adjustment and partial personnel replacement of Tencent's Hunyuan team has sparked discussions on professional social networks and within the AI circle. Many employees were caught off guard, and their astonishment was palpable. This sense of shock is not hard to understand: after all, not long ago, at the release node of HY4 Preview, the team had just gone through a round of high-intensity sprint work followed by a short vacation; what is more, Yao Shunyu, the former core OpenAI researcher who is regarded by the outside world as the new leader of Tencent's AI business, has just joined the company to take charge of foundational model R&D. From the outside perspective, this was supposed to be a moment of gearing up for action and large-scale expansion, so why has it been met with personnel downsizing?
One factual boundary needs to be clarified first: Tencent has never officially defined the changes to the Hunyuan team as "layoffs" or "large-scale personnel optimization". At Tencent's general meeting of shareholders in May 2026, President Martin Lau clearly responded that "there is definitely no large-scale layoff plan, and Tencent is not the same as Silicon Valley companies". What is happening to Hunyuan is not a top-down layoff campaign, but an organizational restructuring characterized by department mergers, reporting line streamlining, and partial team personnel replacement after Yao Shunyu took office. The difference between the two is not a word game, but the key to understanding the direction of Tencent's AI strategy.
If you only focus on the rise and fall of a single team, it is easy to misinterpret it as the setback of a single business. However, from the macro perspective of Tencent's overall AI strategy and the cycle of China's large model industry, this "transition between the old and the new" is not only inevitable, but even a necessary path for Tencent to achieve self-redemption in the deep water zone of AI development.
This is not just a simple personnel change, but a comprehensive shift of Tencent's large model business from the "extensive land grab relying on human sea tactics" to the "saturation attack led by core elite forces".
The accounting period after the fanaticism: squeezing out the "excess water" of large model teams
Looking back at 2023 and 2024, it was the golden age of the entire large model track in China. In order to quickly catch up with the technological wave triggered by ChatGPT, major tech giants launched a cost-agnostic arms race.
At that stage, the expansion of large model teams was extensive and aggressive. Tens of thousands of algorithm engineers, data annotators and evaluation personnel were recruited with high salaries. The industry consensus at that time was "miracles come from sufficient investment". In order to quickly narrow the generational gap with overseas giants, major tech companies needed to build a complete pipeline covering data cleaning, manual labeling, basic fine-tuning and model evaluation in the most efficient way. In this process, the Hunyuan team expanded rapidly, and a large number of junior engineers and middle-level managers were absorbed into the team.
However, by 2026, the situation has undergone a fundamental reversal. The underlying technical paradigm of large models has been relatively stable, and the marginal effect of the Scaling Law has begun to decline. When the parameter scale of models reaches the trillion level, simply relying on stacking computing power and increasing data volume can hardly bring amazing qualitative changes. At the same time, the extreme expansion of computing power costs has forced the CFOs of major tech companies to turn their attention to labor costs.
What is even more critical is that with the improvement of the capabilities of large models themselves and the maturity of the internal tool chain, those "dirty and tiring tasks" that used to require a lot of manual work — such as basic data cleaning, conventional fine-tuning tests, and adaptation of edge scenarios — are being replaced by automation and more advanced models.
As the fanaticism recedes, major tech companies start to calculate their accounts. When the large model business enters the second half from "proving that it can be developed" to "proving that it can make profits", the efficiency pressure is inevitably transmitted to the departments with the highest capital consumption. For major tech giants, foundational model R&D has become an oligopolistic game. Continuing to maintain a huge conventional army not only brings low marginal returns, but also causes severe organizational bloat. Cutting redundant positions in edge business lines and tilting resources to the most core key teams is the only solution to maintain competitiveness.
The arrival of the new leader and the technical rejection reaction
Behind this organizational downsizing, there is a drastic reconstruction of the technical roadmap.
Yao Shunyu's joining is a very rare big move in Tencent's AI personnel layout. This 1998-born graduate from the Yao Class of Tsinghua University and PhD of Princeton University joined OpenAI in 2024 to participate in the R&D of intelligent agent products, and joined Tencent in December 2025 as the chief AI scientist of the "CEO/President Office", reporting directly to Martin Lau. He also serves as the head of the AI Infra Department and the Large Language Model Department. In July 2026, Tencent merged the Large Language Model Department and the Multimodal Model Department into the Foundational Model Department, which is managed uniformly by Yao Shunyu. He coordinates the development of language, multimodal, reinforcement learning and Agent directions, and has become the de facto "top leader" of Hunyuan's technology R&D system. From the outside perspective, he is the "reinforcement" Tencent invited to win a turnaround in the next-generation models (HY4 and subsequent versions). The arrival of a top scientist will inevitably bring cutting-edge technical vision, training methodologies, and new thinking on the underlying architecture.
However, behind the aura of the scientist is the ruthless organizational metabolism.
The establishment of a new technical roadmap often means the end of the old R&D logic. If Yao Shunyu wants to reconstruct the underlying technology stack of Hunyuan, and focus on native multimodality, complex Chain-of-Thought (CoT) reasoning, and more efficient reinforcement learning (RL) frameworks, the knowledge structure of the original team will face severe challenges.
Teams that used to only be responsible for fine-tuning specific business lines, or focused on maintaining the old architecture, can hardly find their place in the new blueprint. The arrival of a new leader is often accompanied by the exit of old members. This is not workplace infighting targeting individuals, but an unavoidable "rejection reaction" in the evolution of technology. If Tencent wants to build a "special forces" that can keep pace with OpenAI and DeepMind, it must streamline the bloated command chain, and concentrate resources in the hands of a small number of top "talented generals", instead of continuing to support a conventional army that relies on numerical advantage to boost ranking scores.
From "competing for benchmark scores" to "competing for implementation": Tencent's strategic pragmatism
Reviewing the evolution of Hunyuan team's positioning within Tencent can help us see the internal logic of this round of adjustment more clearly.
In the first half of the "hundred-model war", Hunyuan's core task was to "ensure that it does not fall behind". But in the second half, Tencent's consideration for AI has fully shifted to application implementation and ROI (Return on Investment).
Recently, Tencent's actions in the AI field are extremely pragmatic. On the B-end, WorkBuddy, relying on the ecological advantages and productization capabilities of WeCom, has established a leading position in the office intelligent agent track. According to data from QuestMobile, the total monthly active users of WorkBuddy across multiple platforms reached 32.353 million in September 2026, ranking first in the typical office Agent category, with a quarter-on-quarter growth rate of 647.1%. On the C-end, although Yuanbao is facing strong suppression from Doubao — Doubao had 419 million monthly active users in September 2026, while Yuanbao had 37.33 million, and the two are no longer in the same magnitude — Tencent did not blindly follow the trend to burn money for user acquisition. Instead, it deeply embedded AI capabilities into national-level applications with huge daily active users such as WeChat, QQ, and Tencent Meeting.
This shift of strategic focus directly determines the talent demand of the Hunyuan team.
In the past, large model R&D was "technology-driven", and everyone competed for benchmark scores and parameters; now, large model R&D is "scenario-driven", and the competition focuses on who can solve specific business pain points with lower reasoning costs. Tencent no longer needs so many people who only know how to write general algorithms, but is in urgent need of compound talents who understand both large models and business scenarios, and can seamlessly integrate AI into the WeCom and WeChat ecosystems. Employees who cannot adapt to this transformation from "technology-oriented" to "business-oriented" naturally become the cost of this round of structural adjustment.
The coming-of-age ceremony for AI of major tech giants: abandoning romance and embracing common sense
This round of adjustment of Tencent's Hunyuan is not only a correction of its own strategy, but also a microcosm of the entire Chinese large model industry moving towards maturity.
Not only Tencent, but also major tech giants including Alibaba, ByteDance and Baidu are carrying out varying degrees of "reducing excess weight and building muscle" for their large model teams after the aggressive expansion in the past two years. In the first half of 2026, all four giants completed the action of "merging scattered model teams into one core department": ByteDance's Seed continued to integrate, Alibaba established the Token Foundry business unit, Tencent formed the Foundational Model Department, and Baidu promoted the BMC integration. The rules of the game of the entire industry have changed: the stage of blindly pursuing the parameter scale of models and the size of teams is over. The focus of competition in the next stage will be the extreme control of reasoning costs, solid data barriers, and killer-level application scenarios.
For those employees who left the Hunyuan team in this round of adjustment, their personal sense of loss is real. But from the perspective of objective industrial development, these engineers with practical experience in major tech companies and experience in thousand-card / ten-thousand-card cluster training, flowing into AI applications across thousands of industries, will in turn accelerate the popularization of technology. They are no longer just "problem-solving experts" who deduce parameters in the cloud, but will become the spark that truly realizes the commercial value of AI technology in the business world.
For Tencent, this pain is inevitable. Large model R&D is no longer a game where you can go online just by writing a few lines of front-end code. It is a marathon focusing on underlying computing power, top talents and ecosystem efficiency. At the critical node when Yao Shunyu takes office and the transition between the old and the new takes place, only by drawing back its fist, and using the most elite team to tackle the most core technologies, can Tencent ensure that it remains in the game in the upcoming decisive battle of AI.
This article is from the WeChat official account "Tech Business", written by Tech Business, and published with authorization from 36Kr.