Quantization and Pruning Methods to Make Your LLM Leaner

A team fine-tunes a model for three weeks, gets the evaluation numbers they wanted, and then tries to actually serve it. The checkpoint alone is 140GB. That single number rules out almost every GPU a normal company has sitting in a rack, forces a rewrite of the deployment plan, and turns what should have been a launch week into a scramble for four A100s nobody budgeted for. That moment is more common than it should be, and it's almost always avoidable. The model didn't need to ship at full precision with every parameter intact. It needed to ship as the leanest version of itself that still does the job, and the two techniques that get you there — quantization and pruning — are neither exotic nor new. They're just underused by teams who assume "make it smaller" means "make it worse." This article walks through what each technique actually does, why skipping them costs real money and real latency, and then gets hands-on with five specific methods people are running in production right now, each with working code you can adapt today. These two get lumped together constantly, and it's worth separating them cleanly before going any further, because they…