PEFT
Also called Parameter-Efficient Fine-Tuning
A family of methods that adapt a model by training a small slice of parameters, not the whole thing.
Think of it like
Like customizing a car with bolt-on accessories instead of rebuilding the engine — cheaper, reversible, and you keep the original.
Example
A team uses PEFT to spin up ten task-specific variants of one base model, each adding only a few megabytes of trained parameters.
How it actually works
PEFT is the umbrella over techniques — LoRA, adapters, prompt tuning — that freeze most of the model and train only a small set of new or selected parameters. This cuts compute, memory, and storage, and makes adaptations portable and stackable. It rarely beats full fine-tuning on the hardest tasks, but it wins overwhelmingly on cost and flexibility, which is why it dominates practical tuning.
For product teams
The category that made custom models routine — small, swappable, cheap adaptations of one shared base.
For engineers
Umbrella for methods training a small parameter subset atop a frozen model (LoRA, adapters, prompt tuning).
Related
- LoRA — Includes LoRA.
- Adapter — Also includes adapters.
- Fine-tuning — Contrast with full fine-tuning.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome