Fine-Tuning
Fine-tuning continues a model's training on your own examples so it adopts a style, format or specialty — it changes the weights, where a prompt merely instructs them.
Take a published model, train it further on a smaller, focused dataset, and you get a new variant that behaves differently by default: answers in your format, speaks your domain's language, follows your house style without being reminded. That weight-level permanence is the difference from prompting, which starts from scratch every conversation.
It earns its cost in narrower cases than most people expect. Instructions and reference documents in the prompt solve most needs; fine-tuning pays off when the same behavior must hold across thousands of runs. Full fine-tuning needs far more memory than inference, but lightweight methods — LoRA and friends — train a small add-on to the weights and run on capable consumer GPUs.
One consequence is visible all over the model hubs: most listed models are community fine-tunes of a handful of originals. The base model still decides the hardware story — a fine-tune of a 70B model is a 70B model.
Where you see it on the radar
The board tracks first-party originals rather than the flood of fine-tuned derivatives — so a restyled variant never masquerades as a new model, and hardware verdicts always describe the base that determines them.
See the originals on the board →Related terms