In this study, we propose the FusionAVP algorithm, aimed at accurately predicting antimicrobial peptides (AVPs). To better represent the features of peptide sequences, we treat the feature representations as two modalities: traditional encoding and large model encoding. Traditional encoding focuses on the local biological information of peptides, while large model encoding captures the global structure and complex patterns. To effectively integrate these two types of features—one focusing on global patterns and the other on local biological information—we developed a multimodal fusion strategy based on the Cross-attention mechanism, which enables efficient integration of information from different modalities, thereby enhancing the accuracy of AVP prediction. To further improve the performance of FusionAVP, we employed contrastive learning to optimize the embeddings generated from both the large model-based and traditional encodings. This optimization process enhanced the model's ability to distinguish between AVPs and non-AVPs. FusionAVP outperformed existing AVP prediction algorithms across multiple evaluation metrics, demonstrating a significant improvement in prediction accuracy. Moreover, the model's performance was validated in real-world applications, where it consistently outperformed traditional peptide identification methods.