[Hy4 preview optimization 'Overthinking': Reduction in task iterations and token consumption] Beating AI Newsflash: Tencent's HunYuan and WorkBuddy have carried out a round of targeted optimizations for Hy4 preview, which has now been fully rolled out on WorkBuddy. According to the official statement, this optimization primarily addresses long deliberations and excessive self-validation in complex tasks, significantly reducing task iterations and input/output token consumption without compromising task effectiveness. Hy4 preview is Tencent's open-source flagship MoE released at the end of August, featuring a total of 770B parameters, 49B active parameters, and support for 1M context length, with a focus on long tasks such as coding, office work, and scientific research. When Tencent released it, they proactively listed 'long deliberations and excessive self-validation in complex tasks' as known issues. [Original link]