Tencent Hunyuan and WorkBuddy gave Hy4 preview a targeted optimization pass, and the change is now fully rolled out in WorkBuddy. The update is aimed at the long reasoning and excessive self-verification Hy4 preview can fall into on complex tasks. The teams say it significantly cuts task turns and input/output token use without sacrificing task quality.
Hy4 preview is Tencent's open-source flagship mixture-of-experts (MoE) model, released in late August. It has 770B total parameters and 49B active parameters, supports a 1M-token context window, and is built around long-running code, office, and research work. Tencent itself flagged overthinking on complex tasks and excessive self-verification as known issues when the model shipped.