@latepostnews: 29-Year-Old Yao Shunyu Takes Over: 300 Days of Reforming Tencent Hunyuan - In 2024, Tencent's high-level recruitment team met Yao Shunyu at a top academic conference. At that time, the young man born in 1997 was still a researcher at OpenAI, and he was introduced to Tencent President Liu Chiping. A year later, he returned to China and became the head of Tencent's large language model...

29-Year-Old Yao Shunyu Takes Over: 300 Days of Reforming Tencent Hunyuan - In 2024, Tencent's high-level recruitment team met Yao Shunyu at a top academic conference. At that time, he was a 1997-born researcher at OpenAI, and was introduced to Tencent President Liu Chiping. A year later, he returned to China and became the head of Tencent's large language model. - Yao Shunyu's fortune lies in that, before his arrival, Tencent's top decision-makers had already realized that AI might be a do-or-die battle, and Tencent was far behind. Ma Huateng said at the 2025 employee meeting: "A year ago we thought we were on board, but then we discovered the boat was leaking." They had cleared obstacles in advance — the key leaders of Hunyuan in various positions had all left, Yao Shunyu could report directly to the Group President's Office, and TEG President Lu Shan was a gentle manager willing to delegate maximum authority. - When he first entered Tencent, Yao Shunyu made a request to President Liu Chiping: from the release of the new model for at least one year, he hoped the President's Office would not look at benchmarks. One year was roughly the time to train two generations of models. "Martin (Liu Chiping) agreed." - When joining Tencent as an advisor, Yao Shunyu had only one task: to diagnose the reasons for Hunyuan's long-term lag. The diagnosis result: "Simply put, almost every link was leaking." Hunyuan overly pursued benchmark scores, putting benchmark training materials into the training set, contaminating the data. The model became very good at exams, but performed poorly in real-world scenarios. At that time, the acceptance line for data labeling accuracy was set at 95%, but the actual rate had long stayed at 60%-70%. The team produced a large amount of unusable data just to meet deadlines, and the algorithm team tacitly allowed this. - Hunyuan was not smooth from the start. When the project was initiated in 2023, the project team didn't even have a single GPU; in the end, they borrowed 2,000 from the advertising department. The lack of cards led to Hunyuan's infrastructure (Infra) being naturally deficient; the system lacked design for large-scale training tasks. The training pipeline was also incomplete; Hunyuan had hardly done any reinforcement learning. Under competitive pressure, the team made two choices. One was to change the architecture, taking a risky route of mixing Transformer and Mamba, a path not yet fully validated in the industry, "but it's a gimmick." Apart from Tencent, few major companies chose it. The other path was to chase benchmarks, reporting scores upward. - Within months after Yao Shunyu arrived, the heads of Hunyuan pre-training, post-training, evaluation, and Infra were all replaced, with newcomers from ByteDance, Kimi, DeepSeek, and Meituan. Hiring no longer looked at background — currently, the head of model architecture at Hunyuan is still a PhD student. ByteDance's Seed recruitment team found that several candidates they wanted had even not been snatched by Seed. The reason: Seed already has high talent density, so young people are more likely to become "cogs"; while Hunyuan is in a rebuilding phase and urgently needs people. - The reforms were not a storm. Yao Shunyu did not mass-fire old employees, but the pressure he brought was more subtle — he often posted papers in group chats to discuss technology, and very few could keep up, sometimes even understanding. "Peer pressure is too high." - The department hierarchy was simplified to only three levels: Yao Shunyu — direction lead — researcher/intern. When not rushing a release, Hunyuan's intensity was not high. On the other hand, Hunyuan also canceled mid-year performance reviews, encouraging a focus on long-term R&D. Occasionally, there were exceptions. "As soon as a group gets a person from Seed, that group quickly starts to voluntarily work harder." - Yao Shunyu used his former employer OpenAI to encourage the team: even today, OpenAI's foundation models have never relied on any mysterious technology. Building large models has no magic, and don't believe others have magic. The real difficulty is getting all the basic, correct things right — doing these is enough to push Hunyuan into China's first tier. - At the end of May 2026, just one month before the official release of Hunyuan Hy3, a batch of submitted data had problems. Yao Shunyu rarely lost his temper, sternly warning the team: "Data is very important. If this happens again, leave immediately." - Hunyuan Hy3 is a large model that Yao Shunyu began training early this year, and it is his first report card since joining Tencent, but internal expectations were not high. Hy3 is not entirely without ambition. According to Yao Shunyu's vision, Hunyuan does not need to beat frontier models like Claude Opus in all abilities. If a model can be as good as or even better than Opus on 90% of daily questions at 1% of Opus's price, it is a better model for most users. Hy4 is already in training, and the Hunyuan team predicts entering the domestic first tier by 2027. - In Yao Shunyu's conception, the model and product should cooperate in a co-design model, i.e., the model and product are developed together from the start. The model can get real scenarios and user feedback from the product in real time, knowing where to improve; the product can also synchronize requirements to the model in a timely manner, without waiting for a general model to be delivered and then adjusting and making do. Yao Shunyu knows his main task at present: to develop good relationships with business units and implement the co-design idea. - Today, the large model industry generally faces monetization challenges, but Tencent's thinking is that it is not necessary to charge external customers; helping business units increase revenue is also the value of Hunyuan. "For example, Honor of Kings earns tens of billions annually. If Hunyuan can help it increase by 1%, that's hundreds of millions." - At Tencent, once a product receives high-level attention, bosses start to give intensive opinions, and people from other supporting departments join in, with more and more participants in decision-making. Last year, Tencent pooled company resources to support Yuanbao, so senior leaders would frequently give very specific suggestions: whether the font size in a certain place is too small, not friendly to older people; why a certain design uses that color. "As soon as the top mentions something, the bottom starts speculating: what does the boss mean, should we change it or not?" - In 2019, Tencent introduced a batch of engineering and data experts from Google, Snapchat, and Uber. Most of them left without significant results. Today, Yao Shunyu has authorization from the President's Office, has carved out a territory, and is the sole decision-maker. But when reforms go deeper, involving larger interests and higher risks, only the President's Office can bear the pressure on his behalf. At that point, the test is not only on him but also on whether the trust between him and the bosses can hold. Reform, at its deepest level, never depends only on the reformer.