Tencent Hy4 Preview is a flagship mixture-of-experts model from Tencent's Hy team, released as an early open-weight preview on August 28, 2026. The full model carries 770 billion parameters but only activates 49 billion per token, drawing on 256 routed experts plus a shared expert across most of its 78 layers and selecting eight routed experts per step. A native multi-token prediction layer is included to accelerate inference. The model is built to handle very long inputs, with a one million token context window that makes it well suited to lengthy coding sessions, multi-document analysis, and other tasks where keeping a large working memory matters.

Tencent positions Hy4 Preview for demanding agentic and analytical use cases, including long-horizon coding, tool-using agents, office analysis, game development, and scientific work. The weights have been released openly, including an FP8 variant for lighter deployments, and access is available through Tencent Cloud TokenHub, OpenRouter, or self-hosting. Because it is still labeled an early preview, prospective users should expect the model to occasionally over-spend effort reasoning or verifying its work, and independent reliability evidence at launch remains limited. Practical fits include enterprise document workflows, code-heavy research assistants, and game or simulation prototypes that benefit from a long context and tool integration.