Tencent quietly releases MixPose Hy3 preview: Yao Shunyu joins as the first flagship model

According to Dynamic Beating monitoring, Tencent has released the latest flagship model Hy3 preview in the MixedElement series, adopting the MoE (Mixture of Experts, where multiple small-scale expert networks are activated on demand) architecture, with a total of 295 billion parameters and 21 billion activation parameters. The model has been launched on the OpenRouter API and HuggingFace demo but has not yet released open-source weights. This is the first flagship model led by Yao Shunyu, former OpenAI researcher and Tencent's Chief AI Scientist, after joining Tencent, and it is the output in less than 3 months since Tencent rebuilt its core model R&D system in February this year.

Tencent positions it as "on par with top domestic models, but still lagging behind cutting-edge models from the United States such as OpenAI and Google DeepMind." The model emphasizes agent capabilities and has been specifically optimized for mainstream agent tools like OpenClaw, supporting a 262.1K context window and a three-level reasoning mode. Currently, it has been deployed in Tencent products such as Yuanbao and CodeBuddy.