Skip to content

Glossary entry

Fine-Tuning

Adapting a pre-trained model with additional domain data so it performs better on a narrower set of tasks. It pays when it lets a smaller, cheaper model clear the quality floor at high volume; otherwise retrieval and better prompts are usually cheaper.

Why it matters

Fine-tuning adds a one-off training cost and a recurring burden of evaluation, hosting and retraining. Whether that pays depends on volume and on what the alternative costs, not on whether the tuned model is better in a demo.

Fine-tuning changes a model's weights using examples from the organisation's own work. Its costs come in three parts:

  • The tuning run, a one-off charge for compute, plus the work of preparing and labelling the data.
  • Hosting, which some providers price above the base model, or which requires dedicated capacity.
  • Upkeep: evaluating the tuned model, and repeating the exercise when the base model is updated or the task changes.

The economic case is usually one of two:

  1. A smaller model that clears the quality floor. If tuning lets a cheaper model do work that would otherwise need a frontier model, the saving per call can repay the tuning at high volume.
  2. Behaviour that prompts cannot reliably produce, such as a strict output format or domain style, where the alternative is costly review and rework.

Where the need is for current or organisation-specific facts, retrieval-augmented generation is usually cheaper and easier to keep up to date than tuning. Compare the options on cost per successful outcome, including review time, over the expected life of the model. See The Economics of Model Selection.

Where this term comes up