Modern deep learning techniques, such as the ET-BERT model, show significant promise in network classification and intrusion detection. However, foundational models built on transformers require substantial resources for training, including pre-training, labeled data for fine-tuning, and memory and computational power for inference. The goal of this project is to investigate the trade-offs between data availability, model size, and accuracy metrics to effectively balance the resources needed in the cloud with those available in local networks and/or log files. By understanding these trade-offs, we aim to optimize the deployment of deep learning models for cybersecurity, making them efficient and feasible for use in router hardware and other local network devices.