%20--%3e%3cdefs%3e%3cstyle%3e%20.st0%20{%20fill:%20%23061b40;%20}%20.st1%20{%20fill:%20%23306af1;%20}%20.st2%20{%20fill:%20%235ce5cf;%20}%20%3c/style%3e%3c/defs%3e%3cg%3e%3cpath%20class='st0'%20d='M55,10.5h9v3h-9v9h12V7.5h-12v3ZM64,19.5h-6v-3h6v3Z'/%3e%3cpolygon%20class='st0'%20points='69%2016.5%2078%2016.5%2078%2019.5%2069%2019.5%2069%2022.5%2081%2022.5%2081%2013.5%2072%2013.5%2072%2010.5%2081%2010.5%2081%207.5%2069%207.5%2069%2016.5'/%3e%3cpolygon%20class='st0'%20points='95%2010.5%2095%207.5%2083%207.5%2083%2022.5%2095%2022.5%2095%2019.5%2086%2019.5%2086%2016.5%2095%2016.5%2095%2013.5%2086%2013.5%2086%2010.5%2095%2010.5'/%3e%3cpath%20class='st0'%20d='M40,1.5v21h11.6l1.4-1.4v-7.6h0c0,0-1.4-1.5-1.4-1.5l1.4-1.4V2.9l-1.4-1.4h-11.6ZM50,19.5h-7v-6h7v6ZM50,10.5h-7v-6h7v6Z'/%3e%3c/g%3e%3cpath%20class='st1'%20d='M23.1,24L14.7,4.8l-4.9,11.2h3.8l-1.8,4H3.3L12.1,0H2C.9,0,0,.9,0,2v20c0,1.1.9,2,2,2h21.1Z'/%3e%3cpath%20class='st2'%20d='M34,0h-16.8l10.6,24h6.2c1.1,0,2-.9,2-2V2C36,.9,35.1,0,34,0ZM32.5,20h-4V4h4v16Z'/%3e%3c/svg%3e)
On July 6, Tencent Hunyuan Hy3 was officially released. Compared to the preview version, it demonstrates significantly stronger intelligence than models of the same size and is comparable to (2-5 times larger) flagship models, with further reduced pricing, and a significant improvement in overall stability and cost-effectiveness. Hy3 has been integrated into multiple businesses such as WorkBuddy/CodeBuddy, Yuanbao, Marvis, and ima. The API is now available on Tencent Cloud TokenHub, and multiple overseas API platforms will be added in the near future.
The model's intelligence has made a comprehensive leap, achieving a qualitative transformation in the practicality of intelligent agents
Hy3 is a model that combines fast and slow thinking, using the MoE architecture, with a total parameter count of 295B and an activated parameter count of 21B, supporting a context length of up to 256K. The Hy3 preview released on April 23 was the first version after the reconstruction of Hunyuan. It achieved a qualitative transformation compared to Hy2 in complex reasoning, instruction following, context learning, code generation, and agent capabilities. Hy3 continues to show a clear and steep growth curve in its abilities, further improving post-training computational scale and data quality and diversity, resulting in a further enhancement in various tasks compared to the Hy3 preview, achieving performance comparable to large-scale flagship models for the first time with a smaller size.
The capabilities of the Hunyuan model are rapidly improving
Hy3 has made particularly significant progress in productivity tasks such as software development, office production, financial modeling, front-end design, and game creation, making it a reliable and cost-effective choice. In a blind test involving 270 experts based on real work scenarios, Hy3 (average score 2.67 / 4) outperformed GLM5.1 (average score 2.51 / 4), especially showing significant advantages in categories such as front-end development, data and storage, and CI/CD.
Proven through real-world user scenarios, significantly enhancing user experience
Hy3 has been tested in the use of global developers and in Tencent's extensive real business scenarios. Since the release of the preview, daily token consumption has increased by 20 times, indicating that the market is beginning to recognize the positioning of high cost-effectiveness and practical models.
As the most popular AI office intelligent agent in China, WorkBuddy has real and complex scenario requirements such as automated script generation and workflow orchestration, which have provided high-value directions for the iteration of the Hunyuan model. Since its release, the number of users who have autonomously selected the Hy3 preview on WorkBuddy has grown six times. In internal evaluations of the WorkBuddy office scenario, the task success rate of Hy3 increased from 72% to 90%, and the average time was shortened by 34%. It makes solving various problems more convenient and efficient.
Yuanbao's dialogue interaction scenarios have also provided valuable feedback for the model. Addressing issues like hallucinations in long text and AI search scenarios, Hy3 has learned to produce reliable outputs under complex evidence through fine-grained data cleaning and training constraints. In evaluations based on real logs, the common sense error rate of Hy3 decreased by half compared to the preview version, and the hallucination rate decreased by more than half. Based on the significantly improved agent capabilities of Hy3, Yuanbao has also launched agent functions. In its internal evaluation, Hy3 has exceeded many domestic excellent models such as GLM 5.1 in both comprehensive office and life service scenarios, and is sufficient to stably support real business processes. Users can input their needs in daily conversations, and Yuanbao can directly execute complex tasks and deliver files such as PPT, Word, Excel, PDF, and HTML, meeting daily office needs, and all for free.
ima evaluated Hy3 based on two core scenarios: online knowledge base Q&A and agent capabilities. In agent tasks, Hy3 performed excellently, with system stability reaching 95.1%. Its tool orchestration capabilities were particularly outstanding, with a significant reduction in ineffective operations such as blind retries and failure to stop. This allows for more accurate planning of complex office tasks, completing them in one step. The knowledge base Q&A scenario also showed significant improvements, with a net increase in reasoning quality of nearly 19%, leading to more systematic thinking and broader information coverage. The structural completeness and usability of long-form writing and solution generation have also significantly improved.
In the core scenarios of file editing/creation, file management, and computer diagnostics and operations for Marvis Agent, Hy3 achieved a task completion rate of 93.7%, an increase of 12.7% compared to the preview version. With six agents working together, the correct task dispatch rate reached 92%, an increase of 13.5% compared to the preview version. Additionally, Hy3 demonstrated significant improvements in complex task execution, multi-step agent workflows, office automation, and reducing hallucinations, capable of stably supporting real business processes and bringing substantial improvements to the execution stability, low latency experience, and cost reduction of Marvis Agent.
Hunyuan continues to support various business operations of WeChat and games. In the AI avatar and customer service special evaluation of WeChat official accounts, Hy3's intent recognition accuracy was improved to 98.94%, and it could reasonably judge even when expressions were incomplete, avoiding overthinking or mechanical templates. In WeChat Reading, Hy3's tag annotation accuracy improved by 14.1% compared to the Hy3 preview, and the overall classification efficiency improved by 8.4%. In the game scenario, after the AI game assistant "Path of Exile: Arrival" launched on WeGame was integrated with Hy3, the comprehensive success rate of multi-round reasoning and tool scheduling was improved to 92%, and the hallucination rate dropped from 4.5% to 2.8%, significantly improving output accuracy and optimizing player experience.
Tencent's highly diverse product matrix provides extensive and real feedback for the optimization of the Hunyuan model, while the progress of the Hunyuan model's capabilities has also fed back to all products, forming a virtuous cycle of transferable and generalized benefits. For example, optimizations in intent understanding and output style in dialogue scenarios can also empower various agent scenarios, and the improvement in agent tool calling capabilities has significantly helped business areas such as dialogue search.
Adhering to a high-cost-performance open strategy, benefiting more users and industries
Hy3 continues to maintain its practical and inclusive model positioning, providing services at a high cost-effectiveness price. The pricing is 1 yuan per million tokens for input, 4 yuan per million tokens for output, and 0.25 yuan per million tokens for input that hits the cache.
In terms of open source, Hy3 uses the Apache 2.0 open-source license, which is highly friendly to commercial use, allowing developers worldwide to download and use it freely for commercial purposes. To further facilitate global developers, Hy3 will be gradually launched on multiple overseas platforms, covering OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, CherryStudio, and simultaneously "day 0" access to the open-source model community Huggingface and Modelscope Mada platform.
From the infrastructure reconstruction in early January 2026, to the release of the Hy3 preview in April, and now the progress of Hy3, Hunyuan has completed the entire model R&D process from underlying reconstruction to product feedback within less than half a year. In the future, Tencent Hunyuan will continue to accelerate technological iterations, continuously explore the upper limits of model intelligence, and focus on building practical application capabilities in real scenarios, transforming leading large model capabilities into accessible productivity across various industries.
Related Links:
Hy3 blogpost: 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