Tencent Hunyuan: Hy4 preview model upgrade, significantly reducing the number of task rounds and input/output Token consumption
2026-09-07 17:26Models🔥 42.2 heat score
1sources
1days unfolding
42.2heat score
3mentions
SummaryAI generated
On September 7, 2026, Tencent Hunyuan team, in collaboration with WorkBuddy, announced that its Hy4 preview model had been fully optimized and launched. The model, released on August 28, had a total parameter size of 770 B, with 49 B of active parameters, and a context length of 1M. To address the issues of prolonged thinking and excessive self-validation in complex tasks, this update significantly reduced the number of task rounds and the consumption of input and output tokens without compromising task effectiveness. The optimization process was monitored through both Bench metrics and manual evaluations. Officials stated that the model has achieved significant expansion in terms of size, context length, and data scale. The combined progress of pre-training and post-training improved its intelligence level significantly, placing it firmly among the top open-source models.
The joint team of Tencent Hunyuan and WorkBuddy announced today that they have made special optimizations to the Hy4 preview model to address issues such as “long thinking under complex tasks and excessive self-verification.” This optimization was fully launched. Through dual monitoring using Bench metrics and manual evaluations, the optimized version significantly reduced the number of task rounds and the consumption of input and output tokens without compromising task effectiveness. The Hy4 preview model released by Tencent on August 28 had a total parameter size of 770 B, activated parameters of 49 B, and a context length of 1 M. According to the official statement, this model has seen significant improvements in terms of size, context length, and data scale. The combined progress of pre-training and post-training has greatly enhanced its intelligence level, placing it firmly among the top open-source models.