There's a natural assumption that more is better when it comes to AI models. More parameters, more capability. The biggest model available must produce the best results. In a research benchmark, that's often true. In production, it's frequently not.
Deploying a large model has real costs. It requires significant compute, which means significant money. It introduces latency, which matters a lot in user-facing applications where people are waiting for a response. It demands infrastructure that many organizations don't have and can't easily build. And it often produces more capability than the task actually needs, which is waste in every direction.
This is why model optimization exists as a discipline. The goal isn't to make models worse. It's to make them leaner, faster, and cheaper without losing the capability that actually matters for the specific job they're doing.
The techniques in this section work in different ways. Knowledge distillation and model distillation train a smaller model to replicate the behavior of a larger one, transferring the large model's knowledge into a form that's cheaper to run. Model compression is the broader category that encompasses several approaches for reducing model size. Model pruning removes weights that contribute little to the output, trimming the model without significantly affecting its performance. Model quantization reduces the numerical precision of the model's weights, which shrinks memory requirements and speeds up inference with surprisingly small accuracy loss.
Together, these form a toolkit for AI model optimization: the practice of finding the smallest, fastest model that still does the job well enough.
The practical implication is worth sitting with. A distilled or quantized model running on modest hardware can outperform a much larger model in a real deployment, because it responds faster, costs less per query, and doesn't fall over under load. Size is a starting point, not a destination.
