Services

Fine-Tuning AI Models

Fine-tuning trains an AI model further on examples from your business until it works in your tone, your formats, and your industry's terminology. We tell you up front whether it pays off for you, and we prove the difference with evaluations: measurements on your own test cases.

What fine-tuning is

An off-the-shelf language model can do a lot, but little of it exactly the way your business needs. Fine-tuning trains the model further on examples from your work: an input and the result you expect for it. A customer request and the reply your best person would have written, for instance. Or an inspection report and the fields from it that belong in your system.

After that, the model writes and decides the way your examples show, without long instructions in every request.

We check with you whether fine-tuning solves your task better than good instructions in the request or RAG, which looks things up in your documents. If it pays off, we prepare the examples, train the model, and prove the result on your test cases.

Fine-tuning or RAG?

The two solve different problems. RAG (retrieval-augmented generation) looks things up in your documents at the moment a question is asked and hands the model the relevant passages to read. The model itself stays unchanged: RAG is not training. Fine-tuning changes the model: how it phrases things, which format it sticks to, how it classifies cases.

That gives a simple rule. For knowledge that changes, such as prices, policies, or product data, RAG is the right tool. For behavior that should stay consistent, use fine-tuning. Often the two work together: RAG supplies the facts, the fine-tuned model puts them in your format. Our guide What is RAG? covers the differences in detail.

When fine-tuning pays off

Fine-tuning pays off when a result stops improving with prompting alone, meaning better instructions and examples in the request. Typical cases:

  • Tone: text that goes to your customers sounds like your company, not like a generic assistant.
  • Fixed formats: the model returns the same structure every time, so your systems can process it without manual fixes.
  • Industry terminology: the terms, abbreviations, and classifications of your field come out right, including where a general model gets them wrong.
  • Speed and cost: a small fine-tuned model handles a narrow task that would otherwise need a large one. It answers faster, costs less per request, and can run on your own servers.

When you do not need it

Fine-tuning is not a way to teach a model your company knowledge. It cannot cite a source for anything it learned in training, and every change to that knowledge would mean training again. You do not need fine-tuning when

  • the model mainly has to know facts from your documents; that is what RAG is for,
  • good instructions and a few examples in the request already give the result you want,
  • there are no examples yet that show what a good result looks like,
  • every answer has to cite its source.

When fine-tuning does not fit

If that describes your case, we tell you before a project starts.

Then we recommend the route that solves your task: better instructions and examples in the request, or RAG, which looks up the answer in your documents and cites the source.

How a fine-tuning project runs

It starts with a measurement, not with training:

  • We agree with you on what a good result looks like and put together test cases with expected answers.
  • We measure how far a model gets with good prompting. If that is enough, the project ends there and you skip the training.
  • If it is not, we prepare examples from your data and train the model.
  • The evaluation compares before and after on the same test cases. You see the difference in numbers before the model goes live.

Your data, your project

We fine-tune on your data only when you ask us to and only for your project; we do not use your data for anything else.

The trained model runs where your data is allowed to be processed: on your own infrastructure, or with a provider that keeps your data in the EU under a data processing agreement. Our local AI page shows how models run in-house.

Frequently asked questions

Does the model know more about our company after fine-tuning?

Not reliably. Fine-tuning shapes how the model writes and decides, but it is not a store for facts. If the assistant needs to know your documents, we connect it to them with RAG, and then it can cite the source too.

How many examples does fine-tuning need?

That depends on the task. Quality matters more than quantity: each example has to show exactly the result you want. A first training run, measured against your test cases, shows how many your task needs.

Which models can be fine-tuned?

Open-weight models (models you can download and run yourself) such as Llama, Mistral, or Qwen, and models from providers that offer fine-tuning through their API. Which one fits depends on where your data is allowed to be processed and how fast and cheap the answers need to be.

Does a fine-tuned model need retraining later?

Only when the behavior you want changes, such as a new format, or when you move to a newer base model. New knowledge needs no training; it comes in through RAG. After every new training run, the test cases from the first project show whether the model got better.

How much does fine-tuning cost?

The price depends on the task, the number and state of your examples, and where the model will run. Because we first measure whether prompting is enough, you pay for training only when the measurement shows prompting is not enough. The Project Wizard on this site gives you a first price range in a few minutes.

Have a project in mind?

Get an instant analysis and quote using our AI-powered Project Wizard.