Models New
Tencent releases Hy4, adding image understanding and computer use, and says open weights will follow in days
Published:
Tencent released Hy4, the latest generation of its large language model, on Wednesday. It says Hy4 understands images for the first time and can operate command-line and graphical software, with special training on tools such as Blender and CAD. Tencent also reports gains in instruction following, long documents and long agent tasks.
The model is live in Tencent products including WorkBuddy, CodeBuddy and Yuanbao, through Tencent Cloud TokenHub and through third-party platforms such as OpenRouter. Pricing is unchanged from the earlier Hy4 Preview at about $0.90 per million input tokens and $2.70 per million output tokens. Tencent says it will publish model weights and system cards in the coming days.
Tencent's own figures include a cut in its internal hallucination rate from 12.95% for Hy4 Preview to 7.2%, and first place among open-source models in a blind test of 579 financial tasks with 213 finance professionals, scoring 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017.
Caveat. The benchmark and hallucination figures are Tencent's own evaluations. At press time the weights had not been published: a check of Hugging Face found no public Hy4 repository.
Sources