Tencent announces Hy4 and promises open weights
Tencent says it is the first Hy model with vision and computer use; as of 9 November the weights and system card, promised 'in the coming days', were not published.
- Models & capabilities
- Open weights & ecosystem
- Minor
Tencent launched Hy4, the full version of the language model it had released as a preview in August, and described it as one of the world’s leading open-source models. The announcement said weights and system cards would be released “in the coming days”. When the Hugging Face organisation page was checked on 9 November, it listed only the earlier Hy4-preview repositories, last updated on 28 August, and no Hy4 repository.
Tencent said Hy4 is its first model with visual understanding and with computer use through command-line and graphical interfaces, and that it was trained on professional software such as Blender and CAD. The release page gives no parameter count. It lists API prices of about US$0.90 per million input tokens and US$2.70 per million output tokens, matching the preview, with a cheaper Hy4 Flash “scheduled for release soon”. WorkBuddy and CodeBuddy, Tencent’s agent products, were free from 4 to 15 November, and its cloud platform offered 50% off standard pricing to the end of the year.
The evaluation claims are Tencent’s. They include an average score of 83.18 on its WorkBuddy evaluation, and a financial-tasks blind test with 213 professionals and 579 tasks in which Hy4 scored 3.04 against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3. Tencent also reported that its hallucination rate fell from 12.95% to 7.2% against the preview. On security, it said the model was trained on defensive data with offensive exploitation material excluded, and that two named security labs, Xuanwu Lab and Keen Security Lab, reported no significant concerns in the areas they evaluated. No independent evaluation was cited.
TechNode reported, on Tencent’s own materials, that heavy demand left many WorkBuddy users at their usage limits on the first day, and that Tencent expanded Hy4 inference capacity tenfold and raised free allowances.
The announcement therefore stands as a commitment rather than a release: whether the weights appear, and under what licence, is still to be seen. It comes in the same week as other Chinese labs published, among them Moonshot’s Kimi K3.1 Pro and Baidu’s ERNIE-5.5-Flash-Base.