Explained simply · Free to read · 4 min read

Fine-tuning explained simply: how it differs from prompting and RAG

Fine-tuning means further training a finished AI model on a smaller collection of your own examples, so that it consistently handles a specific task in a fixed style or format. This changes the model's internal numerical values, unlike prompting, where you only give instructions, and RAG, where the AI looks things up in documents before answering. For new factual knowledge RAG is usually the better route; fine-tuning helps more with consistent behaviour. Companies use it, for example, for uniform summaries or classifications; OpenAI, however, is currently winding down its self-serve platform for it (as of October 2026).

Published 11 October 2026

In short

Fine-tuning: Fine-tuning is the further training of a finished AI model on a smaller collection of your own examples, so that its internal numerical values change and it handles a specific task more reliably in a fixed style or format.^1,4,8; Example: An insurer retrains a language model on a thousand old claims reports, each labelled with the correct category. Afterwards the model sorts new reports consistently, without a long set of instructions having to be included in every prompt.; Also called: model tuning · supervised fine-tuning · retraining a model · custom model training · Fine-Tuning

What fine-tuning is

Large language models are first trained in general on huge amounts of text. Fine-tuning continues this training with a much smaller, specialised dataset. Google describes it like this: the model is retrained on labelled examples, and its weights, meaning its internal numerical values, are adapted to the task.

OpenAI explains the difference with an exam. You can practise the material in class for months until you have internalised it. Or you are allowed to bring the textbook and look things up. Fine-tuning is the practice in class: the model learns from many examples how to do a task. Looking things up in the textbook corresponds to RAG.

Important: the result is a new, adapted model. Giving your own files to ChatGPT projects or Gemini Notebook, or storing standing instructions for an AI, is not fine-tuning. There the model itself stays unchanged; it is simply given extra information.

Prompting, RAG and fine-tuning compared

Prompting: you write a good instruction, possibly with a few examples. IBM calls this the least effort, with no extra computing power at all. According to OpenAI it is the best starting point and is often enough on its own for tasks such as summarising or translating.

RAG: the AI looks things up in a document collection before every answer. OpenAI recommends this when the model lacks knowledge because it is new, out of date or internal to the company. According to IBM, it needs specialists to prepare and connect the data. More on this in our glossary entry on RAG.

Fine-tuning: the model is retrained. OpenAI recommends it when format, tone or approach are not reliably right. It requires labelled examples: OpenAI advises starting with at least 50, and Google gives 100 to 500 as a guide for Gemini. According to IBM, it is the most demanding route, because data preparation and training take a lot of time and computing power.

The approaches can be combined. In an OpenAI example, a model had to correct Icelandic sentences. Quality rose from 62 to 70 points with a few examples in the prompt, and to 87 after fine-tuning on 1,000 examples. Adding RAG on top then lowered the score to 83, because the task was a behaviour problem, not a knowledge problem.

When companies use fine-tuning

Google names typical cases: when answers should always follow a particular structure or fixed format, when the model should learn when to be brief and when to be detailed, or when the specialist language differs strongly from general text. One Google example is chat summaries in a fixed format with personal data removed.

OpenAI adds cost reasons: a fine-tuned model can manage with shorter instructions, and sometimes a smaller, cheaper model can then do the task as well as a large one. According to OpenAI and Google, fine-tuning is also worthwhile when a task is too complicated to describe in a prompt.

But both providers give the same advice: first get as much as possible out of good prompts and testing, then turn to RAG or fine-tuning.

What changed in 2026

OpenAI is winding down its self-serve fine-tuning platform. Since 7 May 2026, organisations that have never used fine-tuning can no longer start training runs. Since 2 July 2026, the same applies to organisations that have not used a fine-tuned model in the past 60 days. From 6 January 2027, even active existing customers will no longer be able to create new training jobs. Fine-tuned models that already exist keep running until the underlying base model is retired.

Mistral also lists its fine-tuning interface in its documentation as deprecated and no longer actively supported. Instead, Mistral offers companies custom model training together with its own experts.

Google continues to offer supervised fine-tuning for several Gemini models through its cloud platform. And anyone using a model with open weights can retrain it themselves on their own hardware, as the Open Source Initiative also points out. What open weights are is explained in our glossary entry on open-weight models.

Our assessment: for private users, fine-tuning plays practically no role. What you need in everyday life can almost always be achieved with good prompts or with your own files via RAG.

Sources

Created with AI assistance and editorially approved before publication.