en.Wedoany.com Reported - On May 7, 2026, Tencent Hunyuan released operational data for its next-generation large language model, Hy3 preview, two weeks after its launch. The total volume of Token calls has reached 10 times that of the previous generation Hy2 model, with particularly significant growth in code and agent-based scenarios. The total growth in applications such as WorkBuddy, CodeBuddy, and Qclaw exceeded 16.5 times. Tencent Hunyuan launched the Hy3 preview at the end of April after completing a reconstruction of its underlying technology, achieving a leap in large-scale usage in just two weeks.

This set of data, disclosed by Tencent Hunyuan through its official WeChat public account on May 7, confirms the product-driven pull effect of the technology reconstruction strategy. Outlets such as Shanghai Securities News and IT Home cited the same core data points in their reports. The head of Tencent Hunyuan stated in an official announcement: "Hy3 preview focuses on practicality. To collect user feedback and understand the model's effectiveness in real-world scenarios, Hy3 preview launched a limited free trial event on OpenRouter from the very beginning, making it convenient for developers to try the model for free. During this process, we saw continuous growth in model call volume and also collected considerable positive and negative feedback, which provides direction for the model's subsequent iterations."

Public data from OpenRouter serves as external validation. In its second week online, Hy3 preview topped both the weekly total Token call volume chart and the market share chart, achieving a "double first," and its call volume also ranked first in the two major scenarios of programming and tool calling. The top five applications by call volume on OpenRouter are all mainstream international agent and code tools, indicating that Hy3 preview has gained high recognition from global developers, and its cross-platform, cross-scenario adaptability has been verified through call volume.

Hy3 preview is the first model launched by Tencent Hunyuan after reconstructing its underlying technology. It adopts a Mixture of Experts architecture integrating fast and slow thinking, with a total parameter count of 295 billion, an activated parameter count of only 21 billion, an activation ratio of approximately 7.1%, and a maximum context window of 256K. The model began training at the end of January 2026, taking less than three months from research and development to launch. Performance tests show that in agent applications like CodeBuddy and WorkBuddy, Hy3 preview's first response speed increased by 54%, average task completion time was reduced by 47%, and the task success rate remained at 99.99%; the success rate of the AI PPT generation feature in Tencent Docs integrated with the model increased by 20% compared to the previous generation. The model already supports integration with mainstream open-source agent frameworks such as OpenClaw, OpenCode, and KiloCode, has been launched on the Tencent Cloud Large Model Service Platform TokenHub, and will continue to serve global developers through cost-effective API access and open-source formats.

It has now been fully integrated into over ten core Tencent products, including Tencent Yuanbao, QQ Browser, WeChat Read, and Tencent Docs, with mainline products such as WeChat Official Accounts and Game for Peace being integrated successively. Tencent's Chief AI Scientist, Yao Shunyu, previously stated publicly during the Hy3 preview launch on April 23 that Hy3 preview is the first step in rebuilding the Hunyuan large model, hoping to obtain real feedback from the community and users through open-sourcing and release to help improve the practicality of the official Hy3 version. He also revealed that Tencent is continuing to expand the scale of pre-training and reinforcement learning to raise the model's intelligence ceiling, continuously improving the model's comprehensive performance in real-world scenarios through deep collaborative design with numerous Tencent products, and beginning to explore distinctive model capabilities.

This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com