What is Fine-Tuning?

You start from a pretrained model and continue training on a curated set of input-output pairs — a few hundred to a few thousand — until it internalises the pattern. It's genuinely good at shaping behaviour: a house writing style, a rigid output format, a classification task where you have labelled history. It's genuinely bad at adding knowledge, which is the most common misconception. Fine-tuning on your documentation does not make a model reliably know your documentation; it makes it sound like your documentation. For facts, retrieve.

The costs are ongoing: you own a dataset, a training run, an evaluation loop, and a new version every time the base model improves — and a fine-tune of last year's model is quickly beaten by this year's with a good prompt. Most agent problems are cheaper to solve elsewhere. Velaris leans on routing to pick a suitable frontier model per task, retrieval through the capability mesh so tools are found rather than memorised, and evals to check the result — which handles the majority of what people reach for fine-tuning to fix.