A machine learning team has trained a large transformer-based text classifier that achieves 94% accuracy on the validation set. When deployed to a real-time SageMaker endpoint, the model exceeds the required p99 latency budget and cannot be served on the available CPU instance family within cost constraints. The team wants to significantly reduce inference latency while retaining as much of the original accuracy as possible, and they are willing to run an additional training phase. Which modeling approach best meets these goals?