Tencent Hunyuan released the official version of its latest-generation large model, Hy4, on November 4. Compared with the preview version launched at the end of August, this release not only significantly expands the scale of pre-training and post-training, but also introduces visual capabilities for the first time—combining perception, reasoning, and tool calling to handle image processing, chart analysis, and mathematical visual reasoning. Even more notably, Hy4 supports "computer operation" for the first time, allowing mixed use of the command line and graphical interface; Tencent also conducted dedicated training for professional software such as Blender and CAD, as well as long-horizon cross-application tasks, to refine planning and execution.

image.png

Five application scenarios, performance beats preview while price stays unchanged

Officially, its use cases are divided into five areas: vision and computer tasks, front-end and game development, finance and office work, scientific research and model development, and long-term personal assistants. Its slogan is exactly "born for real productivity." Hy4 is now available on Tencent Cloud and OpenRouter, and can also be experienced in CodeBuddy/WorkBuddy, Yuanbao, and ima; its performance is clearly stronger than the preview, while the API price remains unchanged. To echo open source, Tencent will release the model weights and system card, and is working with the vLLM and SGLang communities and Inferact to provide ecosystem support; the safety evaluation is also included in the card.

image.png

Acknowledging the gap, flash version already on the way

However, Tencent also candidly acknowledged that although Hy4 is already one of the strongest open-source models, there is still a clear gap between it and top closed-source models, especially on highly difficult test sets such as frontier mathematics, long-horizon development, and model iteration. The team said it will continue to steadily expand training scale to close the gap. At the same time, Hy4 flash, which focuses on cost-effectiveness and inference speed, is also planned to launch soon and will iterate in coordination with products at a higher frequency.