PEFT is Hugging Face's open-source library for parameter-efficient fine-tuning of large pretrained models. Instead of updating every parameter in a model, PEFT methods train a much smaller set of additional or selected parameters, reducing the memory, storage, and compute required to adapt large language and vision models.

The library implements widely used techniques such as LoRA and related adapter-based approaches and integrates closely with the Hugging Face Transformers ecosystem. Small adapter checkpoints can be trained, stored, shared, loaded, and combined with base models, making experimentation with multiple task-specific variants much cheaper than maintaining full model copies.

PEFT is especially useful for researchers and engineers who want to customize large models on limited hardware or maintain many specialized fine-tunes. It has become a common building block in modern LLM training pipelines where full fine-tuning would be unnecessarily expensive.