Multimodal sarcasm detection is of great significance for understanding users’ intentions and emotions. Existing methods suffer from weaknesses in visual modality representation, insufficient granularity in cross-modal fusion, and a lack of semantic reasoning knowledge. This paper proposes a multimodal sarcasm detection model based on cross-modal fine-grained fusion and reasoning chain enhancement, conducting research from three dimensions: modal modeling, knowledge-enhanced reasoning, and semantic fusion. (1) We utilize multimodal large language models to perform deep semantic parsing of visual modality images; (2) based on the long-chain reasoning technology of large language models, we achieve knowledge enhancement under the constraints of semantic and logical consistency; (3) we design an image region-text token-level association mapping to achieve more detailed and accurate cross-modal information fusion. Experimental results on the HFM and SarcNet datasets show that the method proposed in this paper has significant advantages over baseline methods.