In the field of Internet of Things (IoT) security, tasks such as network asset discovery and threat intelligence typically rely on IP addresses to identify target devices. However, the dynamic nature of IP addresses leads to identity distortion, whereas device fingerprinting can extract highly distinctive features of target devices, offers an effective solution to this issue. Existing device fingerprinting methods often suffer from limited applicability and poor generalization capability, making them less effective in large-scale Internet environments. To address these challenges, this paper proposes a novel device fingerprint discovery approach based on fine-tuned large language models (LLMs). By leveraging the comprehensive semantic understanding capabilities of LLMs on protocol text data and incorporating expert knowledge to construct a specialized dataset, we fine-tune the model to effectively capture the contextual features of device fingerprints. Experimental results demonstrate that the fine-tuned model achieves a recall rate of 75.7% and a precision of 76.0% on the test dataset. Moreover, it successfully identifies new fingerprints that were not present in the training samples, further validating the effectiveness of LLMs in capturing device fingerprint characteristics and improving generalization performance. The complete source code is available at https://github.com/iansmith123/DeviceFingerLLM .