Tencent Hunyuan Hy3 Officially Launches with Significantly Enhanced Agent Capabilities

El Ciudadano

Original article: Tencent Hunyuan Hy3 se lanza oficialmente con capacidades de agente notablemente mejoradas

On July 6, 2026, Tencent Hunyuan Hy3 was officially launched. Compared to its preliminary version, it exhibits remarkably superior intelligence compared to models of its size, rivaling flagship models 2 to 5 times larger in parameter scale, all while offering lower costs and substantially improved stability and profitability. Hy3 has already been integrated into multiple Tencent businesses, with its API available on Tencent Cloud TokenHub and various global API platforms being phased in.

The Model Intelligence Takes a Significant Leap Forward, Qualitatively Enhancing Agent Practicality

Hy3 is a model that merges both fast and slow thinking. It employs a MoE architecture with a total of 295 billion parameters and 21 billion activated parameters, supporting a context length of 256K. The Hy3 Preview launched on April 23 marked the first version following the reconstruction of Hunyuan, achieving qualitative advances over Hy2 in complex reasoning, instruction tracking, contextual learning, code generation, and agent capabilities. Hy3 continues this pronounced and clear growth curve in capabilities: by further expanding post-training computational power and improving data quality and diversity, it makes another significant leap beyond Hy3 Preview across various tasks, becoming the first model to rival the performance of large-scale flagship models globally while being smaller in size.

The advancements of Hy3 are particularly noteworthy in productivity tasks such as software development, office productivity, financial modeling, front-end design, and video game development, making it a reliable and high-profitability option. In an internal blind testing of models conducted by 270 experts using real work tasks, Hy3 (average score: 2.67/4) outperformed GLM-5.1 (average score: 2.51/4), showing significant advantages in areas like front-end, data and storage, and CI/CD.

Refined in Large-Scale Real User Scenarios, Significantly Enhancing User Experience

Hy3 has been refined through its use by developers worldwide and in Tencent’s numerous real businesses. Since the Hy3 Preview was put into operation, the average daily token consumption has increased twentyfold, reflecting the growing market recognition of its positioning as a practical and high-profitability model.

As China’s most popular AI office agent, WorkBuddy has complex real-world scenario requirements, such as automated script generation and workflow orchestration, providing valuable direction for the iteration of Hunyuan model capabilities. Since the launch of the Hy3 Preview, the number of WorkBuddy users opting for Hy3 Preview independently has increased sixfold. In internal assessments of WorkBuddy office scenarios for Hy3, task success rates leaped from 72% to 90%, and average time consumption decreased by 34%, facilitating and streamlining the resolution of various issues.

Tencent’s extremely diverse product matrix offers broad and authentic feedback to optimize Hunyuan models, while advancements in the Hunyuan model’s capabilities, in turn, benefit all products, creating a virtuous circle of transferable and generalizable improvements. For example, optimizations in understanding intent and output style in conversational scenarios can also enhance various agent scenarios, while improvements in agents’ tool-calling capabilities significantly benefit businesses like conversational search.

Committed to an Open and Profitable Path to Benefit More Users and Industries

Hy3 maintains a practical and inclusive model positioning, offering services at high profitability prices: 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.25 yuan per million input tokens with cache hit accuracy.

In the open-source domain, Hy3 adopts the commercially favorable Apache 2.0 open-source license, allowing developers worldwide to download and use it commercially for free. To further facilitate its use by global developers, Hy3 will be launched on multiple international platforms such as OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, CherryStudio, and more, and it will also be incorporated into open-source model communities like Hugging Face and ModelScope on «day 0».

From the reconstruction of the infrastructure in late January 2026 to the launch of Hy3 Preview in April and the advancements of Hy3 today, Hunyuan has completed the full cycle of model R&D—from foundational reconstruction to product feedback—in less than six months. Looking ahead, Tencent Hunyuan will continue to accelerate its technological iteration, exploring the maximum limits of model intelligence, and focusing on building applications that can be implemented in real-world scenarios, transforming the leading capacities of large models into inclusive productivity across industries.

Links:

Hy3 Blog Post: https://hy.tencent.com/research/hy3

Hy AI Studio: https://aistudio.tencent.com/

Tencent Cloud: https://console.cloud.tencent.com/tokenhub/models/detail?modelId=hy3®ionId=1

GitHub: https://github.com/Tencent-Hunyuan/Hy3

Hugging Face: https://huggingface.co/tencent/Hy3

ModelScope: https://modelscope.cn/models/Tencent-Hunyuan/Hy3

GitCode: https://ai.gitcode.com/tencent_hunyuan/Hy3

La entrada Tencent Hunyuan Hy3 Officially Launches with Significantly Enhanced Agent Capabilities se publicó primero en El Ciudadano.