Tencent's Hybrid Expert Model Hy3 Preview: A fusion of fast and slow thinking.

Tencent's Hunyuan team has released the Hy3 preview large-scale model, the first achievement after a complete overhaul of the team's pre-training, reinforcement learning, and infrastructure. The model employs a MoE architecture that blends fast and slow thinking, boasting 295B total parameters, 21B activation parameters, and 256K long context, making it Hunyuan's most intelligent model to date. Hy3 preview performs exceptionally well on challenging inference leaderboards such as FrontierScience-Olympiad and IMOAnswerBench, demonstrating significantly improved agent and code capabilities, and can generate complete WeChat mini-programs in a single run.