Hy4 preview

Tencent · open weight · hy4-preview

Hy4 preview is a new-generation Mixture-of-Experts (MoE) flagship model from the Tencent Hy Team (released 2026-08-28). It has 770B total parameters with 49B activated per token, plus a native MTP layer (10B total / 0.7B activated) for speculative decoding. The backbone uses 78 layers (first dense FFN, then 77 MoE layers with 256 routed experts + 1 shared expert, top-8 routing), Gated DeepSeek Sparse Attention with IndexCache, and identity Hyper-Connections, with a 1,048,576-token context window. It is text-in / text-out, Apache 2.0 licensed, and a hybrid-thinking model defaulting to high reasoning effort (also supports no_think). Training targets productivity workloads — software engineering, office/analysis, game development, and scientific research — and is co-designed with Tencent products such as CodeBuddy and WorkBuddy. As an early preview, known limitations include longer-than-necessary reasoning on complex tasks and a tendency to over-verify its own work.