In large-scale network environments, the complexity and volume of network traffic pose significant challenges to real-time threat detection. The research examines how BigBird and Reformer and Linformer and Perceiver NLP architectures boost cybersecurity system performance through improved accuracy and operational speed. A transformer-based hybrid system operates to process lengthy security logs which leads to threat recognition for phishing attacks and ransomware incidents. The model utilizes BigBird sparse attention technology along with Reformer memory efficiency alongside Linformer scalability features to perform testing on the UNSW-NB15 dataset. The proposed detection system obtains 95% accuracy alongside 2% false positives along with 33% improved time efficiency relative to standard models. Temporary dataset UNSW-NB15 requires weighted loss calculations because it possesses class imbalances and synthetic traffic elements that were managed by subsampling techniques. FastAPI provides real-time deployment functionality to the system while its scalability characteristics enable usage in enterprise and cloud security infrastructure applications. The research team plans to expand the architecture design through zero-shot learning methods in order to detect multi-stage and zero-day attacks.