As the best substitute for antibiotics, antimicrobial peptides (AMPs) have important research significance. How to accurately identify AMPs using computational methods has been a key issue in the field of bioinformatics in recent years. However, traditional machine learning methods require autonomous extraction and selection of features from sequence information, resulting in low AMPs identification accuracy. Faced with the above challenges, the deep learning prediction methods based on BERT were proposed and achieved excellent prediction performance. In order to conduct a comprehensive evaluation of existing BERT-based AMP tools and further improve the performance of AMP calculation methods, we compared four existing BERT-based AMP prediction tools in terms of pre-training strategies, word vector embeddings, and prediction performance, and a novel AMP prediction tool is proposed. Experimental results show that compared with existing tools, the model improves on multiple performance evaluation indicators.