What Is Fine-Tuning in AI? A Clear Definition

Fine-tuning is additional training performed on an already-trained model, using a smaller, focused dataset, to change how it behaves rather than what it knows. The model's weights, the internal numbers that determine its output, get updated again, but starting from a capable base instead of from scratch. That distinction, behavior versus knowledge, is the one most explanations skip and the one that actually determines whether fine-tuning is the right tool for a given problem.

What actually happens during fine-tuning

A base model is trained once, at enormous cost, on a broad slice of text. Fine-tuning takes that finished model and continues training it, briefly and cheaply by comparison, on a curated set of examples: input and desired-output pairs that demonstrate the specific behavior wanted. The model's weights shift slightly toward producing outputs that look like the examples, without a full retrain from zero. This is why fine-tuning a large existing model is dramatically cheaper than training one, and why it became practical for individual companies rather than only labs with massive compute budgets.

Two ways to fine-tune

LoRA (low-rank adaptation) and similar parameter-efficient methods freeze the original model and train a small number of additional parameters that steer its behavior, which is why they became the default approach for anyone outside a frontier lab: a fraction of the cost and compute for the large majority of the practical benefit.

Do not confuse fine-tuning with distillation. Fine-tuning adjusts an existing model's own weights to change its behavior. Distillation trains a separate, usually much smaller, model to imitate a larger one's outputs; what is model distillation covers that distinction and when it is the better fit.

What fine-tuning is actually good at

Consistent tone and format is the clearest win: a support bot that should always respond in a specific voice, a code generator that should always follow a team's exact style conventions, a classifier that should sort inputs into a fixed set of categories the same way every time. Fine-tuning teaches a pattern of behavior directly into the model's weights, which tends to be more consistent than repeating the same instructions in every prompt.

It also excels at narrow-domain vocabulary and structure, for example legal, medical, or internal jargon a general model was not trained heavily on, and at compressing a long, complex instruction set into behavior the model just does by default, freeing up prompt space for the actual task.

What fine-tuning is not good at

Adding new factual knowledge is the most common mistake. Fine-tuning on a company's product documentation does not reliably teach a model to answer questions about that documentation accurately, because fine-tuning shapes behavior patterns, not a reliable factual lookup, and a model can still hallucinate details even after being fine-tuned on the correct ones. Retrieval, not fine-tuning, is the right tool for grounding answers in specific facts, which RAG vs fine-tuning vs long context covers in full as a decision, including the cases where using both together is the actual answer.

It is also the wrong tool for anything that changes frequently. Retraining every time a price list or policy updates is slow and expensive compared to retrieval-augmented generation, which just points the model at the current version of the document instead.

The risk worth knowing about: catastrophic forgetting

Training a model further on a narrow dataset can degrade its general abilities elsewhere, a phenomenon researchers call catastrophic forgetting. A model fine-tuned hard on customer support transcripts can get measurably worse at unrelated tasks it previously handled fine, because the training process nudged its weights toward the new pattern at the expense of older ones. This is the main reason parameter-efficient methods like LoRA are usually preferred over full fine-tuning outside of large labs: freezing the original weights and adding a small trainable layer on top limits how much of the base model's general capability gets overwritten.

A concrete example

A company wants a model that drafts customer replies in its specific brand voice (short sentences, no exclamation points, always offers a concrete next step) and correctly cites current order and shipping data. Fine-tuning solves the first half: a few hundred examples of ideal replies teach the model that voice reliably, more reliably than a lengthy style guide pasted into every prompt. It does not solve the second half. The order and shipping data comes from retrieval against the live database at the moment of the reply, because that information changes by the hour and a fine-tuned model would otherwise be reciting whatever data existed when it was last trained.

Frequently asked questions

Do I need to fine-tune a model to get good results?

Usually not as a first step. A well-built prompt with a clear example, covered in how to prompt AI without getting generic answers, solves most consistency problems without any training at all. Fine-tuning earns its cost when prompting genuinely cannot get consistent enough results at the volume you need it.

Can I fine-tune a closed model like GPT or Claude?

It depends on the provider, and access narrowed in 2026. OpenAI wound down self-serve fine-tuning in May 2026: organizations that had not already run a job lost the ability to start new ones, and existing fine-tuned models keep working through the API only until their base model is retired. Anthropic has never exposed fine-tuning through its own public API; the only supported path for Claude is a managed fine-tuning service on Amazon Bedrock, or a custom enterprise engagement for larger accounts. This differs from open-weight fine-tuning, where you download and retrain the actual weights yourself on your own infrastructure, and no provider can wind the option down under you.

How much data do I need to fine-tune a model?

Far less than training from scratch. Useful results for narrow behavior changes have come from as few as a few hundred well-chosen examples, though quality and consistency of the examples matters more than raw quantity.

Does fine-tuning make a model smarter?

No. It makes a model more consistent at a specific behavior, not more capable in general. A fine-tuned model does not reason better or know more than its base model, it just does the specific trained-for task more reliably.