Hunyuan researchers release language model that activates 13 billion parameters

The researchers behind Hunyuan-A13B say they have released the language model as open source. It has 80 billion parameters in total but uses 13 billion for each response, an architecture intended to reduce the computing needed to run it. A second feature adjusts how much reasoning the model uses for routine or complex questions. The technical report describes training on a 20T-token corpus followed by supervised fine-tuning and reinforcement learning. Its authors report competitive results across mathematics, science, programming, language and agent tasks, but the supplied results do not establish a measured deployment cost.