Tencent Hunyuan has introduced and made publicly available the preliminary version of its next-generation language model—Hy4. This model features a total of 770 billion parameters, 49 billion activated parameters, and a context window exceeding 1 million tokens. Thanks to these characteristics, Hy4 demonstrates outstanding performance across various real-world productivity tasks, including coding, office work, and scientific research.
The release of the Hy4 preliminary version marks a significant increase in model size, context length, and training data volume. The combination of improvements in pre-training and post-training has led to a major leap in intelligence level, placing the model in the elite group of open-source models.
Hunyuan continues to collaborate closely with products such as CodeBuddy and WorkBuddy to optimize the user experience in work scenarios. In an internal blind test conducted by Tencent involving 163 experts and 203 engineering tasks, Hy4 preview achieved an average score of 2.99 out of 4. This score slightly surpasses the results of GLM 5.3 (2.92/4.00) and Kimi K3 (2.94/4.00).
Positioned as 'built for productivity,' the model has made substantial progress across a wide range of practical tasks. This success is attributed to high-quality data co-developed by leading Tencent experts in software development, gaming, finance, and security, as well as deep collaboration with products like WorkBuddy.
In software development, the model has strengthened its capabilities to understand, plan, debug, and verify development tasks over a large horizon, while also improving the visual quality and interactivity of frontend development.
In office and analytical scenarios, the model shows a significantly deeper understanding of complex work environments and enhanced financial analysis capabilities. Special emphasis is placed on data analysis and cross-file interaction, supporting the entire workflow—from information processing to creating documents, spreadsheets, and presentations.
Regarding game development, the model can generate a playable prototype directly from a single-sentence prompt and effectively work with game engines. Developers can continuously refine complex game projects through multi-turn interactions.
In scientific research, the model has substantially increased its ability to understand, reason, and solve complex research problems, achieving significant successes in areas such as AI R&D, molecular dynamics simulation, condensed matter physics, and fundamental mathematics.
Notably, Hy4 preview contributed significantly to its own end-to-end development process by participating for the first time in the automated optimization of training methods, data strategies, evaluation systems, and low-level operators. The model proposed approaches, ran experiments, and adjusted them based on the results obtained. The code, logs, and feedback generated by these experiments were then used in subsequent research rounds, forming an initial closed loop of recursive self-improvement.
Furthermore, Hy4 preview independently identified bottlenecks in the inference system and conducted several rounds of optimization in areas such as operator fusion and communication efficiency. These efforts increased end-to-end throughput by 31.8% compared to the baseline, with a stable increase observed across all context lengths and parallelism levels—an early indication of the model's ability to autonomously optimize the inference infrastructure.
Hy4 preview is now available to the public and was launched simultaneously in Tencent's products, including Chinese and international versions of WorkBuddy, CodeBuddy, Yuanbao, and ima. Users can utilize the model directly through these products or access it via API through Tencent Cloud TokenHub and OpenRouter. Hy4 preview maintains an accessible and cost-effective pricing policy: $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cache tokens.
To gather additional user feedback and further improve the performance of WorkBuddy and CodeBuddy, users will be granted free access for a limited two-week period. Since Tencent Hunyuan restructured its infrastructure in February, it releases a major model iteration approximately every two months. The launch of a preliminary version followed by a full release allows Hunyuan to continuously integrate real user feedback into the development process, enabling the model to improve by solving practical problems. The next version of Hy4 will be released soon.