- Published on
Steadily among the top tier of open-source models, Tencent releases Hunyuan Hy4 preview model
- Authors
- Name
- DeepHorus
Tencent officially released its latest generation large language model, Hy4 preview, on August 28th, with a total parameter count of 770 billion and 49 billion active parameters, supporting a context length of up to 1 million tokens. This release marks another significant technological breakthrough for Tencent in the field of Artificial General Intelligence (AGI).
Hy4 preview is the latest iteration of Tencent's Hunyuan series models. Compared to its predecessors, the model has achieved significant expansion in model size, context length, and training data scale. Tencent stated that the synergistic optimization of pre-training and post-training phases has led to another substantial improvement in the model's intelligence level, now firmly placing it in the top tier of global open-source large models.
In terms of specific capability enhancements, Hy4 preview has excelled in multiple real-world productivity tasks by collaborating with top experts from Tencent's internal software engineering, gaming, finance, and security fields to build high-quality data. For example, in the software engineering domain, the model has enhanced its understanding, planning, and debugging capabilities for long-horizon development tasks; in office analytics scenarios, it has significantly improved complex financial analysis and cross-file collaboration abilities.
To validate practical application effectiveness, Tencent organized 163 internal experts to conduct blind tests on three models: Hy4 preview, GLM-5.3, and Kimi K3. The test covered 203 real engineering tasks, and ultimately Hy4 preview achieved an average score of 2.99/4.00, slightly outperforming GLM-5.3 (2.92 points) and Kimi K3 (2.94 points), demonstrating its competitiveness in productivity scenarios.
In terms of scientific research, Hy4 preview has also made important progress. When combined with Tencent's self-developed inference framework Hyra, the model successfully solved a century-old classic geometric problemāthe three-dimensional BlaschkeāLebesgue problemāadvancing the lower bound of the volume from 0.380799 to 0.41104. This achievement brings the Meissner tetrahedron conjecture to within just 2% of its final proof.
Currently, Hy4 preview has been positioned as an early version in the iteration process. Tencent acknowledges that there is still room for improvement in aspects such as long-horizon thinking for complex tasks. The model has been open-sourced through Tencent Cloud's TokenHub and OpenRouter platforms and can be experienced in Tencent's products such as WorkBuddy, CodeBuddy, and Yuanbao.
In its announcement, Tencent emphasized that the decision to release the preview version early aims to accelerate optimization and ultimately launch a more powerful Hy4 official version by widely collecting real feedback. The company will continue to deepen cooperation with internal products and experts, dedicated to enhancing the inclusiveness and performance ceiling of artificial intelligence technology.
The release of Hy4 preview not only consolidates Tencent's leading position in the open-source large model field but also provides a strong footnote to the technological innovation of Chinese tech companies in the global AI race. With the participation of more developers and researchers, the application ecosystem of this model is expected to expand further.
