Fine-tuning

What it means

Picture a customer-service team that wants an AI model to answer questions in the company's tone and format, every time. It could write longer instructions for the model. Or it could fine-tune it: give the model many examples of good questions and good answers and train it a little further on them. IBM puts it simply: "Fine-tuning in machine learning is the process of adapting a pre-trained model for specific tasks or use cases."¹ A pre-trained model has already learned from a very large amount of general material. Google's machine learning glossary calls fine-tuning "a second, task-specific training pass performed on a pre-trained model to refine its parameters for a specific use case."² (Parameters are the millions or billions of internal settings a model learns during training.) Fine-tuning changes those settings. That is the difference between fine-tuning and simply giving a model better instructions or reference documents, which leaves the model itself unchanged.

Where the word comes from

In everyday English, to fine-tune is to make small adjustments to something that already mostly works. AI researchers took the word in that sense. We could not confirm who first used it in machine learning, so we do not name anyone. By 2014 it was standard in research papers. One widely cited study found that a model built from parts of another model kept a benefit "that lingers even after fine-tuning to the target dataset."³ The term moved from images to language in 2018. Jeremy Howard and Sebastian Ruder proposed "Universal Language Model Fine-tuning (ULMFiT), an effective transfer learning method"⁴ (transfer learning means reusing what a model learned on one task for another task). Later that year, Google researchers wrote that their BERT model "can be fine-tuned with just one additional output layer"⁵ for new tasks. Since then, the word has followed today's large general-purpose models from the lab into law. The EU AI Act notes of these models: "These models may be further modified or fine-tuned into new models."⁶

How others define it

• IBM: "Fine-tuning in machine learning is the process of adapting a pre-trained model for specific tasks or use cases."¹
• Google: "A second, task-specific training pass performed on a pre-trained model to refine its parameters for a specific use case."²
• OpenAI: "Fine-tuning lets you take an OpenAI base model, provide the kinds of inputs and outputs you expect in your application…"⁷
• The EU AI Act, on general-purpose models: "These models may be further modified or fine-tuned into new models."⁶

Where they agree and differ

All four agree on the core idea: you start from a model that already exists and adapt it, rather than building one from scratch. They differ in emphasis. Google and IBM describe a technical step. OpenAI describes a product: you bring examples of the inputs and outputs you want. The EU AI Act treats fine-tuning as something that can produce a new model, which matters because new models can carry new legal duties.⁶ For a leader, the useful point is that fine-tuning is not a minor setting. It changes the model your organization depends on.

In the news

This year, OpenAI began winding down its self-serve fine-tuning service. Its developer guide now says: "OpenAI is winding down the fine-tuning platform."⁷ Its deprecations page sets out the stages. From May 7, 2026, "Creating fine-tuning jobs or training is not available to organizations that have not previously run fine-tuning." On January 6, 2027, "Active existing customers will no longer be able to create new fine-tuning jobs on this date."⁸ Models that were already fine-tuned keep working for now: "Inference on fine-tuned models will be disabled only when the underlying base model is deprecated."⁸ (Inference means actually using the model.) We are confident about these dates because they come from OpenAI's own pages. We did not find a reason for the change on those pages, so we do not give one.

The strongest view on the other side is that fine-tuning still pays off. OpenAI's own guide says: "You can train a smaller, cheaper, faster model to excel at a particular task where a larger model is not cost-effective."⁷ That remains true. The news is that one major provider is stepping back from offering the service. It is not proof that the method has stopped working.

What it means for your organization

If your team has fine-tuned a model, you own a customized version of someone else's product, and its life depends on their plans. When the base model is retired, the fine-tuned version goes with it.⁸ Teams that keep their examples, rules, and reference material in a place they control, outside any one model, can carry that work to the next model. This publication calls that place the [knowledge core](/definitions/knowledge-core). Fine-tuning still has a place, especially for narrow, high-volume jobs where cost and speed matter.⁷ But it should be a deliberate choice with an exit plan.

One question to ask your team: If the model we fine-tuned were switched off next year, where are the examples and rules we trained it on, and could we rebuild it on another model?