Hy3 preview's usage volume has increased by more than 10 times.
On May 7th, the first model of Tencent’s Hymir technology after reconfiguration, Hy3 preview, was launched just two weeks ago. The total token invocation volume has increased by more than 10 times compared to the previous version. The code and the performance of the intelligent agent scenario are outstanding. It has ranked first in the total ranking and market share on the OpenRouter platform in the past week. This model adopts a hybrid expert (MoE) architecture that combines fast and slow thinking. It has a total of 295 billion parameters and 2.1 billion activation parameters. It supports a 256K long context window. The main feature is “prioritizing the intelligent agent”. The actual test shows that in the intelligent agent scenario, the first response speed has increased by 54%, the average task completion time has been shortened by 47%, and the task success rate has reached 99.99%. It has been integrated into internal products such as Yuanbao and QQ Browser. In the future, it will provide services to global developers through high-cost-performance APIs, open source, etc
.