Tencent merges 2 AI teams under Yao Shunyu to accelerate R&D

Tencent merged its LLM and multimodal teams into one unit under Yao Shunyu to close the AI gap with ByteDance and Alibaba.

Tencent Holdings Ltd. merged its large language model and multimodal understanding departments into a single foundation model unit, placing all core AI research under Chief Scientist Yao Shunyu to accelerate model development and cut internal friction.

"This consolidation eliminates redundant work between teams that were pursuing the same goal from different angles," a Hunyuan researcher familiar with the restructuring told 36Kr. Yao has been conducting intensive reviews of his teams since taking over the Large Language Model Department earlier this year.

The restructuring follows a series of moves by Yao to centralize AI resources. Tencent disbanded its AI Lab earlier this year, folding all core R&D personnel into the Large Language Model Department. The company also recruited talent from ByteDance's Seed Infra and post-training teams. Former OpenAI researcher Tian Yonglong, a Massachusetts Institute of Technology PhD, joined in early July to lead vision-language model development, replacing Hu Han, who resigned to start his own venture.

The reorganization comes as Tencent faces a computing power disadvantage. The company's 2025 capital expenditure reached 79.2 billion yuan ($11.7 billion), compared with Alibaba Group Holding Ltd.'s 126 billion yuan for the fiscal year ending March 2026. ByteDance Ltd. plans to invest as much as $70 billion in AI infrastructure in 2026, according to Bloomberg. With fewer resources, Tencent must prioritize — and Yao's strategy is clear: concentrate talent and computing power on foundational model R&D.

On July 6, Yao delivered his first major achievement since joining Tencent: Hy3, a medium-sized model capable of competing head-to-head with Zhipu AI's GLM-5.2 and DeepSeek V4 Pro. The launch marked Tencent's return to the top tier of China's AI race after months of organizational upheaval.

Why Multimodal Understanding Took a Back Seat

The decision to fold multimodal understanding into the broader foundation model unit reflects a strategic reassessment of returns. Multimodal understanding technology — image recognition, video captioning — has matured to the point where accuracy rates for text, image and video recognition exceed 85%, according to a multimodal researcher cited by local media. The marginal gains from further investment are limited.

The harder problem is visual reasoning — making models understand the semantic meaning behind images and video — which depends on improving language model reasoning capabilities, not multimodal research alone. Meanwhile, the commercial path for multimodal understanding remains unclear. "No user is willing to pay for image recognition, as there are plenty of free alternative products on the market," a product manager for Tencent's Yuanbao app said. The office scenarios users will pay for — document processing, presentation creation, research report generation — depend on reasoning, coding and agentic capabilities.

The Super Workspace Endgame

Tencent's internal consolidation mirrors a broader industry shift. Over the past month, Tencent, Alibaba and ByteDance have all begun scaling back fragmented AI agent product lines and converging toward unified super-workspace portals, according to Beyond the Layout. Tencent integrated its QClaw product center into WorkBuddy, a product some see as a potential third blockbuster after QQ and WeChat. Alibaba is preparing to launch "Qianwen Office," consolidating three agent products under the Qianwen ecosystem. ByteDance renamed its AI coding product TRAE SOLO to TRAE Work, shifting from a standalone tool to workflow collaboration.

The convergence shows that the "thousand agent war" is giving way to a battle for the enterprise entry point. "Agents won't become the next WeChat, but they very likely could become the new Windows," Beyond the Layout commented.

Tencent shares traded at about 20x forward earnings as of late July, a discount to Alibaba's 22x and a premium to the Hang Seng Index's 10x. The restructuring, if successful, could improve R&D efficiency and reduce the estimated $4 billion-plus annual AI spending burden. But with ByteDance and Alibaba outspending Tencent on infrastructure by wide margins, the risk remains that Hunyuan falls behind in the compute-intensive race to build larger foundation models. Yao's Hy3 launch shows progress, but the next test — whether Tencent can sustain its position through the next model cycle — will determine if the reorganization delivers.

This article is for informational purposes only and does not constitute investment advice.