Model optimization is the practice of making a trained model smaller, faster, or both, without retraining from scratch. It's one of the highest-leverage activities in ML engineering because it directly reduces serving cost and latency, often by an order of magnitude.