Running a trained model to produce outputs; optimized via quantization, batching, and caching.