In recent years, ultrasound, as a low-cost and non-invasive imaging technology, is widely utilized in clinical diagnosis and detection of thyroid nodules. However, accurately identifying thyroid nodules heavily relies on the expertise of radiologists, which can consume substantial manpower and resources. Although deep learning methods have been utilized for thyroid nodule recognition, the majority of current approaches rely predominantly on static ultrasound images for detection. This approach overlooks the information captured in the temporal sequence of frames, presenting challenges in accurately identifying thyroid nodules. Therefore, this article delves into the research of ultrasound videos and a method for automatic identification of thyroid nodules based on multi-frame dynamic prompt information is proposed. The Dynamic Video Prompting Network (DVPN) starts by choosing high-quality key-frames from the adjacent video frames of the current frame to forecast video prompts. Integrated into the patch embedding of the current frame are these video prompts, which serve as updated inputs for video feature extraction. To enhance computational efficiency, a caching mechanism called Relative Position Bias Cache has been introduced. By encoding relative positional data, this mechanism empowers the attention system to effectively capture structured details, thereby enhancing the model’s efficacy. Additionally, a mixed attention approach is employed to boost the extraction of textural information. Furthermore, a Swin Transformer based on the MoE (Mixture of Experts) architecture has been employed in the backbone. Overall, DVPN achieved a mean Average Precision (mAP50) of 74.2%, achieving state-of-the-art performance.