This is an open-source large language model based on the Mixture of Experts (MoE) and Multi-Head Latent Attention (MLA) architectures, which performs exceptionally well in complex tasks such as mathematical reasoning and code generation. The model has a total scale of 671B parameters, but only 37B parameters are activated for each token. That is, not all 'experts' participate in the computation when processing inputs; instead, a subset of experts is selected for processing. By activating only a portion of the parameters (37B), the model can complete computations, significantly reducing the costs of training and inference.