Parameter-efficient fine-tuning covers any method that adapts a pretrained model's behavior for a new task or domain without updating the full set of weights. The motivation is straightforward: full fine-tuning of a modern language model requires storing gradients and optimizer state for every parameter, which for a multi-billion-parameter model demands memory far beyond what most teams have available, and produces a full new copy of the model per task.
Common approaches
- LoRA (Low-Rank Adaptation) — freezes the original weights and injects small trainable low-rank matrices at selected layers. See QLoRA for the quantized variant.
- Adapters — small trainable modules inserted between existing layers, with the original weights kept frozen.
- Prefix/prompt tuning — trainable vectors prepended to the input, with all model weights frozen entirely.
Why this matters beyond memory
Beyond fitting on smaller hardware, PEFT methods produce small, portable artifacts — a LoRA adapter for a 7B model might be a few hundred megabytes rather than tens of gigabytes — which makes it practical to maintain many task-specific or client-specific variants of the same base model without storing a full copy of each.
The trade-off
PEFT methods generally underperform full fine-tuning when the target task requires the model to learn something genuinely absent from its pretraining distribution — a new domain's vocabulary or a structurally different task. They excel at adjusting behavior, style, and task-following within what the base model already knows.