The Transformation Journey of Tencent's Hunyuan Over Three Years

Key Points Since the launch of large model research and development in 2023, Tencent has made large language models a top priority for the first time, with a 27-year-old young man in charge; The next version of Hunyuan will be an Agent model, primarily trained under the leadership of Yao Shunyu; Before Yao Shunyu joined, Hunyuan had two previous heads: Zhang Zhengyou and Jiang Jie. Both had backgrounds mainly in computer vision and big data, rather than natural language processing; In 2022, when ChatGPT was released, Tencent was experiencing multiple lows. One of the lows was in its gaming business; prior to this, Tencent faced a larger crisis around 2020: the "incubator magic" of QQ had failed; With a "formula" and more resources, Tencent may train the next generation of foundational models in a shorter time. But it also means that the sweet period for Yao Shunyu and Tencent is likely only about six months. In late November 2025, university graduate Lin Feng attended a closed-door meeting of Tencent's Qingtian Program in Shenzhen. The event was by invitation only, lasting two days, and included activities such as a cruise tour and a visit to Tencent's headquarters, along with a departmental meeting—Yao Shunyu was present. This meeting lasted about two hours, with Yao Shunyu as the opening speaker. He spoke for about 20 minutes but was ambitious. "He said the goal of Hunyuan is to benchmark against the world's top large models," Lin Feng told Caixin's "New Cortex." Lin Feng was impressed by Yao Shunyu; besides "feeling that he is the manager leading the development progress of the Hunyuan large model," he was also one of the few young executives from Tencent that Lin Feng saw on-site. Yao Shunyu joined Tencent in the second half of 2025, previously working at OpenAI. The Qingtian Program is a special recruitment initiative for outstanding graduates proposed by Tencent in the second half of 2023, benchmarking against ByteDance's Top Seed talent program. Simultaneously with the Qingtian Program, Tencent has been conducting a year-long high-salary talent poaching campaign. Xu Lan, who has long been engaged in foundational model recruitment, told "New Cortex" that one of Tencent's most important targets for poaching is ByteDance. Candidates at the 2-2 level from ByteDance can receive a T12 or T13 level upon moving to Tencent, equivalent to a two-level promotion from their original position at ByteDance. Moreover, salaries can directly double, with some even increasing by 200%. "This kind of market has only appeared in the last month; previously, candidates at ByteDance's 2-2 level would settle for T9 or T10 levels, with salary increases of only 30%," Xu Lan said. Tencent in 2025 is what ByteDance was in 2024. In 2024, ByteDance accelerated the development of foundational models, and its first move was also to poach talent—its main target at that time was Alibaba. A year later, Tencent 'took over' and began to poach talent from ByteDance. Several individuals close to Tencent told "New Layer" that Tencent is expected to start large-scale external recruitment for foundational model-related talent by the end of 2024 and initiate a series of organizational adjustments. The joining of Yao Shunyu in September 2025 accelerated this process. Yao Shunyu's New Policy Initially, external talent joining Tencent's Hunyuan primarily came from Microsoft. Starting in December 2024, core members of the Microsoft open-source model WizardLM team, including Sun Qingfeng, former chief researcher of the visual computing group at Microsoft Research Asia, Hu Han, and project creator Xu Can, successively joined Tencent. Then in August 2025, researcher Tan Xu, who studied voice models at Moon's Dark Side, joined Tencent—prior to joining Moon's Dark Side, he worked at Microsoft Research Asia. "Tencent only considers candidates from the foundational model teams of DeepSeek, Moon's Dark Side, ByteDance, and Alibaba; they do not look at other companies," Chen Lifeng, who is close to Tencent's recruitment, told "New Layer." He mentioned to "New Layer" that in mid-2025, ByteDance had incentivized employees by issuing "Doubao virtual shares," effectively raising salaries for its large model team. However, during this round of equity incentives, some ByteDance Doubao employees took the opportunity to join Tencent's Hunyuan, with original total annual packages of about 2.5 to 3 million yuan at ByteDance, now receiving offers of over 3 million yuan after joining Hunyuan. In December 2025, ByteDance again sent a salary increase email to global employees, raising the overall bonus and salary adjustment investments by 35% and 1.5 times, respectively, to ensure that employee compensation remains competitive and incentive returns are "leading at the top level" in global markets. Under ByteDance's defense, Tencent's talent acquisition plan has also intensified. "Yao Shunyu's joining is a key milestone in Tencent Hunyuan's talent recruitment," Xu Lan said. Before Yao Shunyu's arrival, executive recruitment at Tencent "was a normal talent flow among large companies." After Yao Shunyu joined, Hunyuan not only strengthened its recruitment efforts but also upgraded the confidentiality handling of candidate information. Marking Yao Shunyu's joining, Tencent recruited more talent related to large language models (LLM), whereas prior joiners, including Sun Qingfeng, Hu Han, Xu Can, and Tan Xu, mainly focused on multimodal research. Several recruitment personnel told "New Layer" that Yao Shunyu would serve as an interviewer for LLM-related recruitment, with some candidates personally recommended by him. Among these newly recruited individuals, at least three T12-level talents came from ByteDance's Seed team and DeepSeek, with one responsible for pre-training data work and another collaborating with Yao Shunyu on cutting-edge explorations of large models. Some Hunyuan employees have gained increased confidence due to Yao Shunyu's arrival. "They believe Yao Shunyu might secure more incentives for the team, just like ByteDance Seed did for its team members by issuing Doubao shares," Xu Lan said. Shortly after taking over Tencent Hunyuan's large model, Yao Shunyu had contact and conversations with each team member and participated in team-building and dinners with different research groups Yao Shunyu's actions have gained the trust of some employees. One internal employee believes he is "likable, and working with him gives confidence," while another internal employee thinks he has "excellent management skills." However, some members of the mixed yuan team have begun to worry about being adjusted due to insufficient abilities and have started to actively seek external opportunities. "The number of people leaving the mixed yuan may increase, whether voluntarily or involuntarily," said Xu Lan. The personnel changes are just one of the new policies Yao Shunyu has implemented since joining the mixed yuan. In terms of model strategy, he has also proposed concepts different from those of the previous leaders of the mixed yuan large model. Sources close to Tencent's mixed yuan told "New Cortex" that compared to the previous head, Yao Shunyu places more emphasis on post-training—this was expected by many, as Yao Shunyu's previous work was mainly related to post-training. In addition, he plans to start from the pre-training data part, "retraining the mixed yuan from the data," which is equivalent to going back to the starting point. Lin Feng stated that Yao Shunyu had proposed a clear plan to reduce the "release speed" at the closed-door meeting of the Qingyun Plan in November, "He said that the previous mixed yuan model updated versions internally very quickly, sometimes releasing two versions in a week. The first thing he did when he arrived here was to slow down the release speed and refine the overall model before releasing it," Lin Feng said. Before being poached by Tencent, Yao Shunyu had only worked at OpenAI for a year, deeply involved in projects like Operator and Deep Research. Earlier, he had just graduated with a Ph.D. from Princeton, focusing on natural language processing and reinforcement learning. During his doctoral studies, he had two main research achievements: one is ToT (Tree of Thoughts), a computational framework that allows AI to decompose complex problems into multi-step thinking processes and explore multiple paths to ultimately find the optimal path. The other is ReAct, an algorithm that enables large models to reason and act simultaneously. Both ToT and ReAct can be applied in the post-training phase of models, enhancing the model's multi-step reasoning and hands-on capabilities. ToT and ReAct are present in OpenAI's projects like Operator and Deep Research. In other words, when Tencent poached Yao Shunyu from OpenAI, he had only one year of work experience. However, Tencent likely offered the highest standards in terms of position and salary in the domestic large model talent war. In September 2025, market rumors suggested that Tencent offered over 100 million yuan to poach him, which Tencent later denied. This figure may not be accurate, but Yao Shunyu, born in 1998, is certainly the youngest "number one" in large language models (LLM) among major domestic companies. In December 2025, Tencent announced an organizational restructuring of the mixed yuan, officially appointing Yao Shunyu as the Chief AI Scientist of the "CEO/President's Office," reporting to Tencent President Liu Chiping, while also serving as the head of the AI Infra Department and the Large Language Model Department. A source close to Tencent told "New Cortex" that Yao Shunyu usually works in the Hong Kong office. According to another Tencent insider, Yao Shunyu is the only person listed in the personnel roster of the CEO/President's Office This is the first time Tencent has turned large language models into a top-level project since launching large model research and development in 2023. Moreover, the person in charge of this model has direct and comprehensive management and personnel authority over the team. "Liu Chiping comes from a financial background and does not have a technical background. Yao Shunyu reports directly to him, which makes the efficiency of obtaining resource support even higher," said Qin Yingying, a former employee of Tencent AI Lab, to "New Cortex." Previously, Tencent's self-developed model, Hunyuan, had a long upward reporting chain and a complex downward management system. Hunyuan Follows the Run A series of personnel and organizational changes from the beginning to the end of 2025 indicate that Tencent is starting to feel anxious in the field of large models. The direct reason for this anxiety in the large company is that its self-developed model Hunyuan is not performing well in the competition of large models. "When many companies release new models, the models that are usually compared in benchmarks are basically GPT, Claude, Gemini, Qianwen, Ki mi, and DeepSeek; you won't see Doubao, nor will you see Hunyuan," said an industry insider in large models to "New Cortex." This choice basically represents the industry's recognition of who is in the first tier of models. Ma Huateng stated at the shareholder meeting in May 2023, "For industrial revolution-level opportunities, whether you bring out the light bulb a month earlier (or a month later) is not that important over a long time span." However, large models have now become an "experimental science," where "from infrastructure to attention mechanisms, from parameter quantities to optimization algorithms... every link has multiple solutions, and you need to experiment to know whether it works and at what scale of parameters it works." The aforementioned industry insider said. This experimental nature means that while late starters can refer to mature solutions, they will also lack the experience accumulated from experiments compared to other competitors; the slower the action, the greater the gap and the harder it is to catch up. Tencent is the last major domestic company to launch its self-developed large model. On the last day of November 2022, OpenAI released ChatGPT. Four months later, in March 2023, Baidu released its first-generation large model Wenxin Yiyan; in April of the same year, Alibaba released the Tongyi Qianwen large model; in August of the same year, ByteDance also released its first-generation large model Lark (note: later renamed Doubao large model). Tencent's starting speed is slower than these three companies. In February 2023, Tencent began to form a team called "Hunyuan Assistant," gathering talent from various business groups, with Zhang Zhengyou, then head of Tencent AI Lab, serving as the project leader. Multiple former Tencent employees told "New Cortex" that Tencent's slow start in the large model field is largely related to the internal rhythm of the company. In 2022, when ChatGPT was released, Tencent was experiencing multiple lows. One of the lows was in the gaming business. In 2022, the total number of game licenses approved in China decreased by 32% year-on-year, making it the year with the fewest approvals in past years. Another former Tencent employee told "New Cortex" that gaming is one of Tencent's most profitable businesses, and due to the sharp reduction in licenses, the uncertainty of gaming revenue increased, leading the company to implement "cost reduction and efficiency enhancement" that year, even requiring Tencent TEG (Technology Engineering Group), a technical middle platform, to be self-sufficient Later, the AI Lab led by Zhang Zhengyou, who was in charge of the development of the Hunyuan model, was under TEG. "At that time, the AI Lab did not encourage researchers to publish papers. Employees had to rely on university budgets to attend academic conferences," he told "New Cortex." Qin Yingying, a former employee of Tencent AI Lab, stated that when the AI Lab was established in 2016, its main task was still basic research. However, around 2018, the AI Lab was split into two parts, with only a small number of people continuing to engage in AI basic research, while the majority had to start serving the company's gaming and advertising businesses. Moreover, this former AI Lab employee mentioned that before the gaming business fell into crisis, Tencent faced a larger crisis around 2020: the "incubator magic" of QQ had failed. Before 2020, many of Tencent's new businesses and products—such as QQ Music and Tencent Video—were successfully incubated through QQ. At that time, QQ played a role as a distribution channel and entry point—similar to the role Douyin plays in the distribution of Doubao today. However, around 2020, many of Tencent's new businesses struggled to replicate past success paths, such as Tencent E-commerce, the Toutiao competitor TianTian KuaiBao, and Tencent Weishi, which aimed to compete with Douyin and Kuaishou, all of which did not achieve the expected success. In 2020, Tencent's other major distribution channel—WeChat—had reached 1.225 billion monthly active users, becoming a super entry point in the mobile internet era, but it has not yet demonstrated the magic of incubating new products like QQ did back in the day. Among the many new businesses, the only one that Tencent's senior management considered to have potential was the Video Account. The aforementioned former Tencent employee stated that in 2022, Tencent's founder, chairman, and CEO Ma Huateng mentioned at an internal employee meeting that "the most outstanding business of the WeChat business group is the Video Account, which is basically the hope of the entire company." By 2023, Ma Huateng again stated at the shareholders' meeting that "AI is an opportunity that comes once in several hundred years, similar to the Industrial Revolution," but throughout 2023 and 2024, this former Tencent employee claimed that he had not heard the company emphasize large models from an overall strategic perspective; the focus remained on the Video Account. According to the aforementioned former Tencent AI Lab employee, in 2023, if different business units within Tencent wanted to train models, they needed to apply for budgets and project approvals from the company's general office. "At that time, the cost of training models was still very high, with each trial and error costing hundreds of thousands of dollars. Even if the training was successful, what was the model's purpose? At that time, no one could clearly explain it," this former Tencent AI Lab employee stated. Difficult to Hold the Number One Position The three-year arms race for large models has made the industry realize that large models are not only an experimental science but have also created an unprecedented development paradigm: they cannot be completed primarily through top-down explicit planning and division of labor, as was the case in the industrial era, along with extensive collaboration among engineers. Unlike the development model that relies on various technical backbones, large model development is a system engineering process. Without a comprehensive experimental conception in the mind of the number one position regarding the model from pre-training data to pre-training architecture and algorithms, post-training paths, and infrastructure levels, a model cannot emerge from the brute-force development of multiple technical backbones The aforementioned Tencent AI Lab employee told "New Layer" that for a long time during the three years after the launch of large model research and development, Tencent did not match a suitable technical leader for the Hunyuan team. The WeChat team also trained a model but later abandoned it. The aforementioned Tencent AI Lab employee informed "New Layer" that before Yao Shunyu joined, Tencent Hunyuan had two general leaders: Zhang Zhengyou and Jiang Jie. Zhang Zhengyou's research direction is computer vision, and his invented method for calibrating flat-panel cameras is widely adopted globally, known as the "Zhang Method." However, the core model in large models is LLM, and the core of LLM is natural language processing, not image processing. "After managing for nearly half a year, Zhang Zhengyou did not deliver satisfactory results. Jiang Jie then took over the Hunyuan project because he believed that large models could serve advertising placement," Qin Yingying said. Jiang Jie joined Tencent in 2012 and is the Vice President of Tencent's Corporate Development Group (CDG) and Technology Engineering Group (TEG), responsible for the technical management of Tencent's advertising platform products. Under Jiang Jie's leadership, Liu Wei and Wang Di both briefly managed Hunyuan directly for a period. "Liu Wei was more like a coordinator, while Wang Di was more like an executor," Qin Yingying told "New Layer." However, both Liu Wei and Wang Di's technical backgrounds are not deeply related to natural language processing. Liu Wei's core work direction is computer vision, and after leaving Tencent at the end of 2024, he founded a company called ReBirth, focusing mainly on video generation. Wang Di joined Tencent through campus recruitment in 2008 and was primarily responsible for data and search advertising algorithms in TEG before leading the Hunyuan LLM model. In November 2024, Jiang Jie also stopped managing the LLM part of the Hunyuan project and was only responsible for the Hunyuan multimodal model. The aforementioned Tencent AI Lab employee stated that at that time, the entire LLM team was waiting for a so-called "senior talent from Microsoft," but this Microsoft executive was never found. It wasn't until September 2025 that Yao Shunyu arrived. Before Yao Shunyu took charge of the Hunyuan LLM model and reported directly to Tencent President Liu Chiping, the head of Hunyuan's LLM was a middle-level role that found it difficult to push work both upward and downward. Qin Yingying told "New Layer" that during the leadership of Zhang Zhengyou and Jiang Jie, there was a group leader for the Hunyuan LLM model, but his rank was not high, and there were at least three levels separating him from Liu Chiping, including LLM technical leader Wang Di, Hunyuan large model leader Jiang Jie, and TEG President Lu Shan. This structure made it impossible for him to communicate directly with Liu Chiping and obtain resource support like Yao Shunyu could. In terms of management, this group leader also needed to cross multiple departments within TEG to reach the project-related members. Established in 2023, Hunyuan is a virtual team, with members coming from the AI Lab and even multiple departments within TEG. Qin Yingying told "New Layer" that departments such as the Machine Learning Department and Data Platform Department belong to the old organizational structure of TEG, not the Hunyuan team. Employees reassigned to participate in the Hunyuan project also had to serve the AI needs of different departments and businesses within Tencent "The past six months at Hunyuan have been quite chaotic, with responsibilities and authorities not clearly defined. Neither Zhang Zhengyou nor Jiang Jie has been able to truly integrate the team," she said. It wasn't until April 2025 that Tencent organizationally equipped a team similar to Byte's Seed, establishing the Large Language Model Department and the Multimodal Model Department within the TEG system. After announcing Yao Shunyu's joining in December, Tencent further adjusted its organizational structure, newly establishing the AI Infra Department, AI Data Department, and Data Computing Platform Department, for the first time aligning the organizational structure with the R&D needs of the large model era. Several recruiters and Tencent employees referred to "New Cortex," stating that Tencent AI Lab has gradually been weakened, with several members having left, and AI Lab Deputy Director Yu Dong will also leave at the end of January. Xu Lan told "New Cortex" that the current focus of AI Lab is exploratory research and publishing papers externally. This means that the AI Lab, established by Tencent in 2016, has effectively become a subordinate department of Hunyuan, rather than its upper management institution. "New Cortex" has learned that with Yu Dong's departure, the research directions of voice and natural language processing in AI Lab will also be cut. The remaining two directions are the multimodal understanding direction led by Hu Han and a unified direction for multimodal understanding and generation led by a former member from Byte's Seed. It wasn't until the end of 2025 that Hunyuan transformed into a structure closer to Byte's Seed, while similar organizational adjustments were initiated by Byte in the second half of 2023. This late but significant adjustment means that the obstacles facing Yao Shunyu, the 27-year-old leader of the large language model, have basically been cleared by Tencent. Accelerating Theory and Sweet Period Entering the third year of the large model arms race, the optimal path for latecomers to catch up with early movers has become very clear: to quickly deliver models that meet market expectations, it is best to directly imitate proven effective mature practices, following the successful paths others have taken, rather than thinking about how to innovate and overtake on a curve. The best example illustrating this success principle is Google. Before 2025, the phrase "getting up early but arriving late" described Baidu in China, while in Silicon Valley it referred to Google—who invented the foundational architecture of large models, Transformer, but has always been considered behind OpenAI's GPT and Anthropic's Claude in the global model competition since the release of ChatGPT. Industry insiders in the large model field told "New Cortex" that before the release of Gemini 2.0, Google had consistently adhered to the Encoder+Decoder architecture (note: encoder + decoder, the original form of the Transformer architecture invented by Google in 2017). When releasing Gemini 1.5 in February 2024, its model paper still emphasized why this architecture was superior, but by the end of 2024, when releasing Gemini 2.0, Google no longer promoted how the Encoder+Decoder architecture was full of potential Instead, it began to adopt a Decoder only architecture like OpenAI. After no longer stubbornly insisting on its self-developed architecture, Google quickly re-entered the industry spotlight with Gemini 2.5 released in March 2025, and the release of Gemini 3.0 in November directly triggered a "red alert" at OpenAI to cope with the traffic pressure brought by Gemini. According to November data from SimilarWeb, the average user stay time on Gemini has surpassed that of ChatGPT. An industry insider told "New Cortex" that the V3 and R1 models released and open-sourced by DeepSeek in December 2024 and February 2025, respectively, also played the role that the Decoder only architecture of GPT once played—showing what a model similar to GPT-4 and OpenAI o1 might look like. Among them, the V3 model provided the industry with a "recipe" to train a model equivalent to GPT-4—indicating that to achieve similar model performance, the model needs to be a MoE (Mixture of Experts) architecture, rather than a dense model; the R1 model released more than a month later provided a "recipe" for a reasoning model similar to OpenAI o1—indicating that reinforcement learning can be directly achieved through imitation learning on corpora with reasoning steps. "Previously, everyone actually didn't know how GPT-4 was trained, nor how o1 was trained," said the insider. Throughout 2024, the main goal of domestic large model companies is to "catch up with GPT-4," but they don't know how to catch up, as OpenAI has not publicly disclosed its core model architecture and algorithms, nor published technical reports since GPT 3.5. The release and open-sourcing of DeepSeek's V3 and R1 models accelerated the global launch of models similar to GPT-4 and OpenAI o1. Hiring a researcher who has directly participated in the post-training of models at OpenAI and has been exposed to its pre-training as the head of the LLM model is also a form of acceleration for Tencent's Hunyuan. Two months after Yao Shunyu joined Tencent, in early December 2025, Tencent released the Hunyuan 2.0 model, with a total parameter count of 406 billion. Its English name was updated to HY2.0, seen as a move to emphasize the global market. Some speculate that this model is the result of Yao Shunyu's work after joining. However, based on the industry cycle for training a generation of models, Hunyuan 2.0 is unlikely to be a model trained from scratch by Yao Shunyu. "The relationship between model parameter count and model capability is a necessary but not sufficient one; a large parameter count does not necessarily mean strong model performance, but a small parameter count certainly limits the model's potential," said the insider. Yao Shunyu is likely to train a model with a larger parameter count, which is estimated to take about half a year—under the conditions of having a "recipe" and more resources, the time may be shorter. This means that Yao Shunyu and Tencent's sweet period is about half a year, and this 27-year-old young manager needs to deliver some tangible results to secure the top position in large models at Tencent, which has 110,000 employees In April 2025, Yao Shunyu wrote an article titled "The Second Half" on his personal blog, announcing that large models have sufficient reasoning capabilities, and the competition for large models has completed the pre-training battle of the first half, entering the second half focused on further training and Agent development. However, at Tencent, he has to re-engage in the first half. A person close to Tencent Hunyuan told "New Cortex" that the next version of Hunyuan will be an Agent model, primarily trained under the leadership of Yao Shunyu. Risk Warning and Disclaimer The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at one's own risk