Tencent's hybrid model upgrade reduces task rounds and token usage

On September 7, 2026, Tencent's Hybrid team, in collaboration with WorkBuddy, announced the full rollout of an optimized version of the Hy4 preview model. The update addresses issues such as 'long thinking' and 'over-self-validation' in complex tasks. The improvements, verified through Bench metrics and human evaluation, significantly reduced task rounds and input/output token consumption without compromising performance. The Hy4 preview model was launched on August 28, 2026, with a total parameter count of 770 B, active parameters of 49 B, and a context length of 1M. Tencent stated that the model's expansion in size, context length, and data scale, combined with advancements in pre-training and post-training, led to a substantial increase in intelligence, positioning it among the top open-source models.