Exclusive | Hu Han, Head of Hunyuan Multimodal Understanding, Resigns to Launch His Own Startup, and the Original Team Is Expected to Focus on World Models

Text | Xinyu ZHOU

Editor | Yuxin ZHANG

Exclusive from "Intelligent Emergence": Recently, Han HU, who leads multimodal understanding at Tencent's Hunyuan, has tendered his resignation.

Prior to this, he served as a principal researcher in the Visual Computing Group at Microsoft Research Asia. After joining Tencent in early 2025, he took charge of research on vision large models. In subsequent organizational adjustments, he joined the "Frontier" advanced technology research group under the Large Language Model Department, leading research related to multimodal understanding and reporting to Shunyu YAO.

It is understood that Han HU has also undertaken R&D work on world models.

Meanwhile, Shunyu YAO, head of Tencent's Large Language Model Department, has recently been conducting intensive reviews of his teams, the research group Han HU previously worked in may focus on cutting-edge research into world models.

As of press time, Tencent has not issued a response to the above information.

Large Language Models Remain the Top Priority for Hunyuan

Since mid-2026, personnel adjustments around the Hunyuan ecosystem have been ongoing. In early July, Yonglong TIAN, a former OpenAI researcher and PhD from the Massachusetts Institute of Technology, joined the Large Language Model Department — we have learned that he will replace Han HU to lead R&D in the vision-language model (VLM) direction, reporting to Shunyu YAO.

Since Shunyu YAO joined Tencent, filling capability gaps and rapidly elevating the foundational model capabilities of Hunyuan into the first tier has become the core priority of Tencent AI.

Restructuring resource allocation to concentrate talent and computing power on foundational model R&D under the Large Language Model Department is one of the core measures Shunyu YAO has implemented since taking office.

For example, after Tencent disbanded AI Lab, all core R&D personnel were integrated into the Large Language Model Department. The recruitment intensity for large language model talent has also increased following Shunyu YAO's arrival, including several professionals from ByteDance's Seed Infra team and post-training team.

Later, at the Tencent Cloud AI Industry Application Conference on June 5th, David WALLER, Senior Executive Vice President of Tencent Group and CEO of Cloud and Smart Industries Group, in a sense, posed a pointed question to Shunyu YAO on behalf of the public: Many people say Tencent is lagging behind in AI, do you think we are really behind?

But in fact, "Shunyu is more anxious than Tencent's senior management," a researcher from Hunyuan told "Intelligent Emergence". He mentioned that Shunyu YAO has repeatedly lobbied Tencent's executives to secure more computing power resources for Hunyuan.

Another detail is that "LatePost" once reported that one month before the launch of Hy3, a batch of data encountered issues. Shunyu YAO used unusually strict wording to warn the team: "Data is of utmost importance. If such a situation occurs again, the responsible personnel will be dismissed directly."

Reassessing the Returns on Multimodal Understanding

When none of Tencent's AI business lines have secured a leading position, for Tencent to achieve a breakthrough, it is necessary to reassess the returns of various AI technology directions.

On one hand, the multimodal understanding research that Han HU engaged in has become increasingly mature, and the marginal returns from continuing to invest resources in this field have become very limited.

"In multimodal understanding research, the recognition accuracy for text, images, and videos can generally reach over 85%," a multimodal researcher told us, "The real challenge lies in visual reasoning — enabling models to understand the semantic meaning behind images and videos. This goal cannot be achieved solely through multimodal research; instead, it requires enhancing the reasoning capabilities of language models."

As the technological dividend gradually fades, the commercial benefits brought by multimodal understanding have also remained ambiguous.

In the past, multimodal research once occupied a core position in Tencent's AI system. The key reason was that computer vision technology could be strongly integrated with Tencent's "cash cow" businesses such as advertising, gaming, and video-audio services, generating additional revenue for the company.

However, compared to generative technologies, scenarios corresponding to multimodal understanding, such as "image recognition" and "generating captions for images", cannot be directly commercialized.

"No user is willing to pay for image recognition, as there are plenty of free alternative products available on the market," a product manager of Yuanbao mentioned. What users are truly willing to pay for are office scenarios such as document processing, PPT creation, and research report generation — which rely on the model's capabilities in reasoning, coding, and agentic operations.

On the other hand, Tencent's computing power resources are limited, and it does not have an advantage compared to ByteDance and Alibaba.

Tencent's 2025 financial report shows that the company's annual capital expenditure reached 79.2 billion yuan; in contrast, as of March 2026, Alibaba's fiscal year capital expenditure hit 126 billion yuan; and according to a Bloomberg report, ByteDance plans to increase its maximum investment in AI infrastructure to 700 billion US dollars in 2026.

With limited resources, the Hunyuan team inevitably faces a situation where demand outstrips supply. At this point, pausing multimodal understanding research with limited technological dividends to focus on more future-oriented cutting-edge research directions is a tradeoff Tencent made after evaluating potential returns.

On July 6th, Shunyu YAO delivered his first official achievement after joining Tencent: the large model Hy3. Among medium-sized models with similar parameter counts, Hy3 is already capable of competing head-to-head with GLM-5.2 and DeepSeek V4 Pro.

Within Tencent AI's own development framework, the organizational restructuring, reconstruction of the infrastructure system (covering data, Infra, and evaluation), and the strategy of refocusing on foundational models implemented since Shunyu YAO's arrival have earned Hunyuan a seat at the table of the AI Tier 1 camp in just half a year.

(Zhaofeng LI, author of "Intelligent Emergence", also contributed to this report.)