CyberSecQwen-4B: A Small, Specialized AI Model for Defensive Cybersecurity

CyberSecQwen-4B is a 4-billion-parameter language model fine-tuned for narrow cybersecurity tasks like CWE classification and CTI Q&A. Developed by lablab-ai in the AMD Developer Hackathon, it runs on a single 12 GB consumer GPU, avoiding the cost and data exposure of hosted APIs. The model retains 97.3% of Cisco's 8B Foundation-Sec-Instruct's CTI-RCM accuracy while exceeding its CTI-MCQ score by 8.7 points. Training used MITRE/NVD CVE-to-CWE mappings and synthetic Q&A, deduplicated against CTI-Bench. The base model is Qwen3-4B-Instruct-2507, fine-tuned on a single AMD MI300X via ROCm 7. A companion 2B model, Gemma4Defense-2B, shows similar performance. The model is Apache 2.0 licensed and available on Hugging Face. It is designed for local, air-gapped environments and explicitly not for generating exploit code or making autonomous security decisions.