Fine-tuning takes a general-purpose foundation model (GPT, Llama, Mistral) and continues training on a smaller domain-specific dataset — legal contracts, medical notes, brand voice examples, structured JSON outputs — so the model performs better on that task without needing long prompts every time.
Most production LLM applications do not need fine-tuning. Prompt engineering and RAG solve the majority of cases faster and cheaper. Fine-tuning becomes worthwhile when: you need highly consistent output format, you have a large body of examples, latency matters (fine-tuned smaller models beat prompted larger ones), or you have proprietary data that is expensive to fit into every prompt.
Modern fine-tuning uses methods like LoRA (low-rank adaptation) that update only a small fraction of the model's parameters — dramatically cheaper than full fine-tuning.