Tencent released and open-sourced the preview version of its Hunyuan Hy4 large model on August 28, 2026, joining a growing wave of Chinese frontier models made available to the global developer community.
Hy4 preview features 770 billion total parameters with 49 billion activated parameters, a mixture-of-experts architecture with 163 experts, and a one-million-token context window. In an internal blind evaluation across 203 tasks, the model scored 2.99, ahead of comparable releases including GLM-5.3 and Kimi K3 in the same benchmark.
Tencent said Hy4 preview has already been deployed in first-wave applications including WorkBuddy, Yuanbao and ima, and that open-sourcing the model allows developers worldwide to self-host, fine-tune and deploy it, supporting the continued growth of China's open-source AI ecosystem.
Image: Efficiency improvement of AI-related computer chips. Source: Wikimedia Commons (CC0).
