This paper proposes a Multi-Model Contrastive Decoding algorithm and uses a dynamic hallucination detection mechanism in the generation phase to significantly reduce the model's hallucinations.