[REVISION HISTORY]

Tencent Hunyuan Hy4 model development

1 recorded revision of this situation.

What changed

2026-11-04 06:58 UTC → 2026-11-05 00:48 UTC · added ~~removed~~

Tencent announced the upcoming release of its Hy4 and Hy4 flash large language models at the 2026 Tencent Global Digital Ecosystem Conference. The Hy4 model was designed to improve reasoning efficiency and long-range productivity, while the Hy4 flash model focused on cost-efficiency through deep co-design between the model and its products. Following the announcement, Tencent officially released the Hy4 version of its Hunyuan large language model. This version expanded pre-training and post-training scales compared to the previous preview and introduced visual understanding capabilities. The model is capable of handling fine-grained images, complex charts, and mathematical problems involving visual data. It also supports computer operations via command-line and graphical user interfaces, with specialized training for professional software like Blender and CAD. Tencent reported that Hy4 showed improvements in instruction following, long-text comprehension, and deep research. In testing involving 213 financial sector experts, the model reportedly ranked first among open-source models. The model is currently available on Tencent Cloud and various Tencent products, with plans to open-source the model weights in the near future. Technical specifications for the launched Hy4 model reveal it is based on a Mixture-of-Experts architecture with 770 billion total parameters, marking a significant upgrade from the Hy3 predecessor. The model is integrated into Tencent products such as WorkBuddy, CodeBuddy, and Tencent Cloud. Tencent has expressed a commitment to responsible open-source development, focusing on cybersecurity and biological safety, and plans to release the model weights and system cards shortly.

Versions

Select a version to see what changed compared to the one before it.