Deploying transformer models on edge devices. Comparing INT8 quantization with structured pruning for model compression. Current models are 1.2GB, need to reduce to <200MB without significant accuracy loss.