Tencent Releases Hunyuan Hy4 Preview, a 770B Parameter Open-Source Model
Tencent has officially released the Hunyuan Hy4 preview, an open-source model featuring 770 billion total parameters, 49 billion active parameters, and a 1-million-token context window. The model is now available on platforms like HuggingFace and GitHub, and has been integrated into Tencent's productivity tools such as CodeBuddy. This release strengthens Tencent's position in the open-source LLM landscape, offering competitive performance against other leading Chinese models like GLM-5.3 and Kimi K3. Additionally, its collaboration with the Hyra research agent demonstrates the practical utility of LLMs in solving complex, long-standing scientific and mathematical problems. In blind tests on 203 engineering tasks, Hy4 preview scored 2.99/4.00, slightly outperforming GLM-5.3 and Kimi K3. However, Tencent noted that as an early preview, the model still exhibits limitations such as excessive self-verification and prolonged thinking times on complex tasks.
## BACKGROUND
Mixture of Experts (MoE) architectures use a routing mechanism to activate only a subset of parameters (active parameters) for each token, reducing computational costs during inference while maintaining a large total capacity. Hyra (Hunyuan Research Agent) is Tencent's recursive self-improving agent designed for performance-oriented research and engineering tasks.