Abstract: Cybersecurity knowledge graph is an important means to conduct deep correlation analysis on unstructured cyber threat intelligence data. Use of large language model (LLM) can overcome the limitations of traditional cybersecurity named entity recognition methods, such as lack of understanding of text context and difficulty in capturing long-distance dependencies. However, the extraction of cybersecurity entities by a single LLM will lead to poor adaptability to the domain outside the model's capability boundary, insufficient generalization ability on out-of-distribution (OOD) data, and cognitive limitations of a single model. To solve these problems, a multi-model collaborative recognition method that integrates fine-tuning and linking was proposed. Low-Rank Adaptation of large language models (LoRA) fine-tuning was performed on a base LLM using a cybersecurity threat intelligence corpus, and the model's understanding of the cybersecurity field was enhanced through parameter optimization; LLM collaboration adopted static voting and dynamic meta-learner linking. The static mechanism integrated prediction results of multiple models through voting, and the dynamic mechanism designed the feature space adaptation function to quantify the model’s capability boundary. Dynamic mapping of input data features to model performance was realized through meta-learners such as random forests. Experimental results show that the F1 score of the proposed method reaches 0.91, which is 6.73% higher than the semantic enhancement network method based on multiple features, providing an efficient and scalable solution for named entity recognition in cybersecurity threat intelligence analysis.