姚顺雨，完成腾讯混元组织大一统

Zhidongxi Author | Li Shuiqing Editor | Yunpeng Zhidongxi reported on July 24 that on July 23, Tencent announced that it would merge the Hunyuan Large Language Model Department and the Multimodal Model Department to form a Basic Model Department, which will be unified and responsible for Yao Shunyu, Tencent's chief AI scientist. This merger means that the large language and multi-modal dual-team parallel R&D model implemented by Tencent Hunyuan since April 2025 has officially ended. Tencent stated that this merger is to further improve model development and collaboration efficiency, and explore the intelligent upper limit of the full-modal model. Yao Shunyu joined Tencent from OpenAI in September 2025 in a low-key manner. In December, he served as Tencent's chief AI scientist, coordinating the research and development of the Hunyuan large language model. In the nearly one year since joining the company, he has pushed Hunyuan to complete a comprehensive reconstruction of the underlying infrastructure. Currently, Hunyuan has formed a complete model matrix based on language models, covering images, videos, speech, 3D generation, etc. Hy3, a large language model officially released and open sourced in early July, adopts MoE architecture, with a total parameter of 295B and an activation parameter of 21B. Compared with the preview version released in April, the agent and code capabilities are improved by 20-30%. It is also leading in performance evaluation at the same parameter level and can rival flagship models with 2 to 5 times the parameter scale. Hy3 has been connected to many businesses such as WorkBuddy/CodeBuddy, Yuanbao, Marvis, ima, etc., and the API is simultaneously launched on Tencent Cloud TokenHub and multiple overseas platforms. In the first week of its release, the total call volume of Hy3 increased by more than 68 times compared with the previous generation Hy2; among users of WorkBuddy’s self-selected models, 60% chose Hy3. In terms of multi-modal models, Tencent released and open sourced the Hunyuan 3D World Model 2.0 (Hy World 2.0) in April. It supports multiple types of input such as text, pictures, and videos. It can automatically generate, reconstruct, and simulate 3D worlds. It is also compatible with the export of 3D assets in multiple formats and can be seamlessly connected with existing game workflows. The number of community downloads has exceeded 3 million. In addition, the mixed element image model 3.0 in LMA