Skip to content
Xavi Creus

AI

Fine-tuning

Fine-tuning is the process of further training a pre-trained AI model on your own examples so it adopts a specific style, format or specialised behaviour.

Definition

Fine-tuning is the process of taking a pre-trained AI model and training it further on a smaller, specific set of examples so it learns a particular style, format, vocabulary or behaviour. Instead of building a model from scratch, you start from one that already understands language and nudge it toward your task. Fine-tuning changes the model's weights, which is what distinguishes it from prompting, where the model stays the same and only the input changes.

In business, fine-tuning is useful when you need consistent output at high volume: classifying support tickets into your exact categories, writing in a strict house style, extracting fields from your specific document layouts, or making a small cheap model perform like a large one on a narrow task. It requires hundreds to thousands of quality examples, a testing set, and someone to maintain the model as your data changes.

For most companies in 2026, fine-tuning is the third option, not the first. Better prompting and retrieval-augmented generation solve most problems without it, and frontier models follow instructions well enough that style fine-tuning is rarely needed. The misconception is that fine-tuning teaches a model new facts. It mostly teaches form and behaviour. For up-to-date knowledge, retrieval is more reliable and easier to update.

In practice

A fintech fine-tunes a small open-weight model on 5,000 labelled transactions so it categorises expenses in its own taxonomy with high accuracy, at a fraction of the cost per call of a frontier model.

Why it matters

Fine-tuning can cut cost per task dramatically once a use case is stable and high-volume, but it adds maintenance. Ask your team whether prompting and retrieval have been exhausted first.

Frequently asked questions

What is the difference between fine-tuning and RAG?
Fine-tuning changes the model itself by training it on your examples, which is good for style, format and specialised behaviour. Retrieval-augmented generation (RAG) leaves the model untouched and feeds it relevant documents at query time, which is better for facts that change. Many systems use both.
How much data do I need to fine-tune a model?
For behaviour or format, a few hundred high-quality examples can be enough; for classification with many categories, a few thousand. Quality and consistency matter far more than volume. Noisy or contradictory examples make the model worse, not better.

Need this explained for your company?

One hour with me is usually enough to turn the vocabulary into a decision.

Book a session