Tencent announced the new flagship of the Hy (formerly Hunyuan) series, Tencent Hy4, on November 4. According to the company's official statement, the model has significantly improved in following instructions, understanding long texts and executing multi-step tasks compared to the previous version Hy4 Preview. In addition, Hy4 can understand images and use a computer on its own for the first time in the series.
Tencent plans to release the system board containing model weights and security evaluations in the coming days. For this reason, Hy4 can only be tried via API and Tencent applications for now. Hy4 Preview, released at the end of August, can be downloaded on Hugging Face with the Apache 2.0 license. According to the model card, there are a total of 770 billion parameters in the preview version, and 49 billion of them are activated for each token.
What can Tencent Hy4 do?
Thanks to its new visual understanding capability, Hy4 can process fine details in images, analyze complex graphics, and solve mathematical problems involving images. On the other hand, the model can work on the computer by working both on the command line and in graphical interfaces. Tencent also says it specially trained the model with professional software such as Blender and CAD, increasing its ability to plan long-term tasks.
The company also underlines that it developed the model together with its own products. For example, Hy4 can track chats, emails, calendars and meetings by connecting to WeCom through the office assistant WorkBuddy. On the gaming side, deep integration with Unreal Engine, Unity and Godot stands out. According to Tencent's internal measurements, Hy4 performed 19.1 percent better in the GameDev test and 9.8 percent better in the GameLogic test compared to the Hy4 Preview.
Similarly, Tencent has also optimized Hy4 for financial analysis. The company's blind evaluation with Tencent Financial Technology involved 213 finance professionals and tested the model in 579 real-life tasks. In this test, Hy4 was ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68) with 3.04 points. Additionally, Tencent announced that, according to its own evaluations, the hallucination rate decreased from 12.95 percent to 7.2 percent.
These figures are all based on Tencent's own measurements, though, and independent tests have yet to be published. Moreover, the company admits on the Hy4 research page that there is still a significant difference between it and the most powerful closed models. According to Tencent, this difference is especially seen in areas such as research-level mathematics, long-term coding and advanced visual question answering.
Where can Hy4 be used, what is its price?
Hy4 is currently available on Tencent platforms such as Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao and ima. In addition, the model's API can also be accessed through third-party platforms such as OpenRouter, Hermes and OpenCode. The company also states that integration work is continuing in products such as QQ, Honor of Kings and Tencent News.
There is no change on the price side. According to Tencent, Hy4 is offered at the same price as Hy4 Preview: approximately $0.90 for million input tokens, $2.70 for million output tokens, and $0.045 for million cached input tokens. In addition, the model will be available at TokenHub with a 50 percent discount until December 31, 2026. WorkBuddy and CodeBuddy will be free between 4 and 15 November.
Finally, Tencent announced that it will soon release the lower-cost Hy4 Flash model. Once the weights are shared, Hy4 will enter into direct competition with Chinese open models such as DeepSeek, Kimi and GLM. The video world model Hy-Nomas, which the company shared as open source last month, also showed Tencent's claim in this field.
