The proliferation of cyber threats has created a deluge of unstructured cyber threat intelligence (CTI), making manual analysis impractical and overwhelming. Knowledge graphs (KGs) offer a powerful solution for structuring this data, but their construction is bottlenecked by the foundational task of named entity recognition (NER). While state-of-the-art NER often relies on massive general-purpose or resource-intensive domain-specific models, this paper explores a novel alternative that leverages a lightweight, code-centric language model, DeepSeek Coder 1.3B. We hypothesize that its unique pre-training on a 2-trillion-token corpus, heavily skewed towards source code (87%), provides a strong inductive bias for recognizing the quasi-syntactic nature of cybersecurity entities. We fine-tune DeepSeek Coder on the CyberNER dataset, a public corpus derived from security blogs and annotated with 10 distinct entity types using the BIO tagging scheme to test our hypothesis. This research evaluates the viability of using code-specialized models as an efficient and effective method for populating cybersecurity knowledge graphs, addressing a critical challenge in automating threat intelligence analysis.