Fine-tuning: Fine-tuning is extra training that adjusts an existing AI model's weights using a smaller, targeted set of examples so it performs better on a specific task, style, or domain. It changes the model itself, unlike prompting, which only changes the instructions.
A general model has already learned language from a huge corpus. Fine-tuning continues that training on your data: a few hundred to a few thousand examples of inputs paired with the outputs you want. The model's weights shift slightly toward your pattern, so it produces that kind of output without being told each time.
It is best for things that are hard to capture in a prompt: a very particular writing voice, a classification scheme with subtle boundaries, a strict output format that must be right every time, or a specialized vocabulary. Major API providers offer fine-tuning services, and open-weights models can be fine-tuned by a technical team.
Fine-tuning is not how you teach a model new facts. It does not give the model your knowledge base; for that, use retrieval-augmented generation. It also costs time and data preparation, and a fine-tuned model can drift out of date as the base models improve. Most teams should exhaust prompting and few-shot examples before considering it.
Example at work
An insurance company handles thousands of claim notes a day and needs each one tagged with one of forty internal loss codes. Prompting gets most of them right but stumbles on the edge cases. The data team fine-tunes a model on ten thousand historically coded notes, and the tagged results become consistent enough to automate the routine cases.
Why it matters
Vendors sometimes pitch fine-tuning as the way to make AI 'know your business.' Understanding what it actually does lets you ask the right question: is our problem about style and format (fine-tune), about access to our documents (retrieval), or about clearer instructions (prompting)?
Related terms
- Pre-trainingPre-training is the first and largest phase of building an AI model, in which it learns general patterns from a massive dataset, such as predicting the next word across trillions of words of text. Later phases then shape it into a useful assistant.
- Few-shot promptingFew-shot prompting is including a handful of examples of the input and output you want inside your prompt so the AI model can copy the pattern. It is the fastest way to get consistent formatting, tone, and judgment calls.
- Retrieval-augmented generation (RAG)Retrieval-augmented generation (RAG) is a technique in which an AI system first searches a set of documents for relevant passages and then gives those passages to a language model to answer from. It lets an assistant use your own, current information.
- Model weightsModel weights are the numbers inside a trained neural network that determine how it responds to input. They are what training produces and what a company ships or protects; releasing them publicly is what makes a model 'open weights.'
- Training dataTraining data is the collection of examples an AI model learns from, such as web pages, books, code, images, or labeled records. What is in that data, and what is missing, shapes everything the model knows, how it writes, and which mistakes it makes.