Fine-tuning
Adapting a pre-trained LLM by continuing training on a smaller, task-specific dataset.
Fine-tuning takes a foundation model and continues training it on data specific to your task or domain. The model adjusts its weights to be better at that specific thing — generating in your brand voice, classifying support tickets in your taxonomy, writing code in your codebase's style.
For most operators in 2026, fine-tuning is overkill. Frontier models with good prompts and retrieval handle the vast majority of business tasks better than fine-tuned smaller models do, with no training overhead. Fine-tune only when: prompting plus retrieval has hit its ceiling, you have thousands of high-quality training examples, and the latency or cost benefit of a smaller fine-tuned model justifies the engineering investment.
Where fine-tuning genuinely shines: very narrow tasks (entity extraction, classification), tasks where you need consistent output structure under heavy load, and tasks involving proprietary data the foundation models haven't seen.