Full fine-tuning updates every weight in a model, which needs a lot of GPU memory and leaves you with a full-size copy per task. PEFT trains only a small set of new or selected parameters while most of the original model’s weights stay frozen. With a common add-on method, a bank could keep one base model and store a small add-on each for fraud notes, customer email and policy questions, loading whichever one a request needs.
The add-ons come in a few families. Adapters insert small trainable layers inside the network; in the original adapter paper they came within 0.4% of full fine-tuning on a language benchmark while adding 3.6% parameters per task (Houlsby et al.). LoRA trains small low-rank matrices alongside the frozen weights and is now the most widely used method. Prompt tuning learns a few “soft prompt” vectors that are fed in ahead of the input; they are numbers learned by training, not words a person writes (Lester et al.). Don’t confuse prompt tuning with tweaking the wording of a prompt, which is prompt engineering. Hugging Face’s PEFT library implements these and many other variants (Hugging Face PEFT).
PEFT matters because it makes customization affordable on modest hardware, and it pairs well with quantization to shrink memory further. It is not always enough. In episode 49 at 11:25, Intercom’s Fergal Reid says his team started out experimenting with LoRA and “parameter efficient fine tuning methods like that” but now mostly runs full supervised fine-tuning on large GPUs, his account of what worked for their scale. Galileo sponsored episode 49 (sponsor read at 28:52). In episode 17 at 5:33, IBM’s Maryam Ashoori places PEFT in a range of options from prompt engineering and RAG to full fine-tuning and alignment tuning, depending on customer needs. Galileo CTO Atindriyo Sanyal co-hosts episode 17 (0:20), recorded while Galileo sponsored the show.
Editorially: try PEFT before full fine-tuning, and judge both on the same held-out evaluation, including checks that general skills did not regress.
Hear it from the guest
“Full fine tuning and parameter efficient fine tuning or even alignment tuning”
“Parameter efficient fine tuning methods like that”
Quotes lightly edited to remove filler words.
Sources
- Hugging Face: PEFT documentation — States that PEFT methods fine-tune only a small number of (extra) parameters, cutting compute and storage while yielding performance comparable to full fine-tuning; lists adapters, LoRA, prefix and prompt tuning; checked October 24, 2026.
- Houlsby et al.: Parameter-Efficient Transfer Learning for NLP — Introduces adapter modules; on GLUE they came within 0.4% of full fine-tuning while adding 3.6% parameters per task; checked October 24, 2026.
- Lester et al.: The Power of Scale for Parameter-Efficient Prompt Tuning — Introduces prompt tuning, which learns soft prompts by backpropagation to condition a frozen model, and finds the gap with full model tuning closes as models grow; checked October 24, 2026.
- Chain of Thought episode 17 at 5:33 — IBM's Maryam Ashoori lists parameter-efficient fine-tuning among the customization options enterprise customers need.
- Chain of Thought episode 49 at 11:25 — Intercom's Fergal Reid says the team started with LoRA-style parameter-efficient methods and moved mostly to full supervised fine-tuning.
Go deeper
- Should you use prompting, RAG, or fine-tuning to customize an AI model? AI, decoded · Fine-Tuning vs. RAG vs. Prompting: How to Choose
- Can switching to an open model actually cut your AI costs? AI, decoded · When an Open Model Beats a Frontier API on Cost
- Li and Liang: Prefix-Tuning paper
A related method that learns continuous prefix vectors for a frozen model; useful for comparing prompt-style PEFT methods.
