Fine-tuning
Fine-tuning is further training of an already-trained model on your own examples so that it adapts to a task, a format or a voice, producing a new set of weights that then has to be served somewhere.
Full fine-tuning updates every parameter and produces a complete new checkpoint. Parameter-efficient methods, of which LoRA is the common one, train a small adapter alongside frozen weights instead — far cheaper to train, small enough to store many of, and swappable at serve time. Almost everything described commercially as fine-tuning today is the second kind.
The training run is usually the smallest of the three costs. Building and cleaning a dataset that is actually representative is the largest, and serving the result is the one that never stops: a fine-tuned model is your model, and it does not share capacity with anyone else's traffic.
It earns its place on a narrow, stable, high-volume task — particularly where a small fine-tuned model can replace a large prompted one, or where an output format has to be exact. It does not earn its place on anything a better prompt or a retrieval step would fix, because every change to the task re-runs the whole cycle.
This service does not offer fine-tuning. Several providers on the comparison pages do, and if training and serving on one platform is what you need, that is a real reason to use one of them instead.
What it costs you
Fine-tuning trades a variable cost for a fixed one. Prompting costs tokens on every request forever; fine-tuning costs a dataset, a training run and then a serving arrangement that is charged whether or not traffic arrives. That arithmetic only works above a volume threshold, and the threshold moves every time per-token prices fall — which has been the direction of travel throughout the open-weight era. Most teams that consider fine-tuning to save money would save more, sooner, by shortening a system prompt.
Related terms
- DistillationDistillation is the training of a small model to reproduce the outputs of a larger one, so that the smaller student performs close to the teacher across a particular range of work while costing a fraction as much to serve.
- Model weightsModel weights are the numeric parameters a training run produces — the entire learned content of a model — and a published checkpoint of them, together with the code and configuration needed to run it, is enough to reproduce that model's behaviour anywhere the hardware exists.
- Open-weight modelAn open-weight model is one whose trained parameters are published under a licence permitting others to download, run and serve them, which is what makes it possible for anyone other than the original lab to offer that model as an API.
Go deeper
Every rate quoted above is published in full on the pricing page, alongside the model publisher’s own official rate.