Last Friday, Tencent released Hunyuan 2.0, its latest self-developed large AI model, offered in both Think and Instruct versions. The model is built on a mixture-of-experts (MoE) architecture with 406 billion parameters and 32 billion active parameters, and supports an ultra-long 256K context window. Compared with the previous generation, Hunyuan 2.0 Think features major improvements in pretraining data and reinforcement learning strategies, delivering strong performance in complex reasoning tasks across mathematics, science, and coding. Hunyuan 2.0 is already integrated into Tencent’s AI-native products, including Yuanbao and ima, and is available via Tencent Cloud’s APIs, allowing users to access its reasoning and creation capabilities. [IThome, in Chinese]
Related
