Hy4 preview

Hy4 preview targets long-running productivity work. The open-weight Mixture-of-Experts checkpoint activates 49 billion of 770 billion parameters per token and extends the context to 1,048,576 tokens. This checkpoint remains an early release.

Tencent describes Hy4 as an expansion in model size, context length, and training data over Hy3 preview. Tencent publishes vLLM and SGLang recipes. Reasoning defaults to high, while no_think switches to direct responses. Longer-than-necessary reasoning and over-verification are listed as known issues.