Han Hu, head of multimodal understanding at Tencent’s Hunyuan unit, has reportedly tendered his resignation.
Before joining Tencent in early 2025, Hu was a principal researcher in the visual computing group at Microsoft Research Asia. At Tencent, he led work on vision foundation models. Following an organizational reshuffle, he moved to a “frontier” advanced technology research group within the large language model (LLM) department, where he oversaw multimodal understanding research and reported to Yao Shunyu. Hu also worked on world models.
Yao, who heads Tencent’s LLM department, has been reviewing teams across the organization. The research group Hu previously belonged to may shift its focus toward frontier research on world models.
Personnel changes across Hunyuan have continued into the second half of 2026. In early July, former OpenAI researcher Tian Yonglong joined the LLM department. According to 36Kr, Tian will succeed Hu as head of R&D for vision-language models and report to Yao.
Since Yao joined Tencent, the company has focused on closing capability gaps and moving Hunyuan’s foundation models closer to the industry’s leading systems.
Yao has sought to reallocate resources and concentrate talent and computing capacity on foundation model R&D led by the LLM department. After Tencent dissolved its AI Lab, its core R&D personnel were transferred to the department. Tencent also increased its recruitment of LLM specialists, including several hires from ByteDance’s Seed infrastructure and post-training teams.
At a Tencent Cloud conference on June 5, Dowson Tong, Tencent’s senior executive vice president and CEO of its Cloud and Smart Industries Group, asked Yao a question that reflected outside criticism of the company’s progress:
“Many people say Tencent has been slow in AI. Do you think we really have been?”
“Shunyu is more eager to move faster than Tencent’s senior management,” a Hunyuan researcher told 36Kr. The researcher said Yao had repeatedly pressed Tencent executives to allocate more computing resources to Hunyuan.
LatePost reported that a batch of submitted data caused problems in late May, about a month before Hy3’s formal release, directly affecting the model’s performance. Yao told the team: “Data is very important. If this happens again, you are out,” according to a person familiar with the matter.
Tencent has yet to establish a clear lead across its AI initiatives. To gain ground, it will have to decide which areas are most likely to generate technical and commercial returns.
One area under scrutiny is multimodal understanding, where further technical gains may be harder to achieve.
“In multimodal understanding research, accuracy rates for text, image, and video recognition can generally reach more than 85%,” a multimodal AI researcher told 36Kr. “The difficult part is visual reasoning, which means enabling a model to understand what images and videos actually mean. Multimodal research alone is not enough to accomplish that. The reasoning capabilities of the language model also need to improve.”
As technical gains become more incremental, the commercial returns from multimodal understanding also remain uncertain.
Multimodal research once held a central position within Tencent’s AI organization, in part because computer vision technology could be integrated into the company’s advertising, gaming, and digital entertainment businesses.
Compared with generative AI, however, applications such as image recognition and captioning can be difficult to monetize directly.
“No user is willing to pay for image recognition when there are so many free alternatives on the market,” a Tencent Yuanbao product manager said. “What users are actually willing to pay for are workplace applications such as document processing, PowerPoint creation, and research report generation. Those applications depend on a model’s reasoning, coding, and agentic capabilities.”
Tencent’s disclosed capital expenditure also trails that of some rivals, although overall spending is an imperfect proxy for AI computing capacity.
Tencent’s 2025 financial report showed that its full-year capital expenditure totaled RMB 79.2 billion (USD 11.7 billion). Alibaba’s capital expenditure for the fiscal year ended March 2026 reached RMB 126 billion (USD 18.6 billion). Bloomberg reported that ByteDance was considering capital expenditure of as much as USD 70 billion in 2026 as it expanded its data centers and other AI infrastructure.
With resources constrained, Hunyuan must allocate computing capacity among competing projects. Against that backdrop, the apparent shift away from multimodal understanding suggests Tencent is prioritizing frontier areas that it believes offer greater long-term potential.
On July 6, Tencent released the Hy3 LLM, Yao’s first major model launch since joining the company. Tencent’s benchmarks suggest that Hy3 approaches the performance of midsize models with comparable parameter counts, including Z.ai’s GLM-5.2 and DeepSeek V4 Pro.
By Tencent’s own benchmarks, six months of organizational restructuring, infrastructure rebuilding, and renewed focus on foundation models may have narrowed the gap with leading AI developers.
KrASIA features translated and adapted content that was originally published by 36Kr. This article was written by Zhou Xinyu for 36Kr.
Note: RMB figures are converted to USD at rates of RMB 6.78 = USD 1 based on estimates as of July 27, 2026, unless otherwise stated. USD conversions are presented for ease of reference and may not fully match prevailing exchange rates.