Full Fine-tuning vs PEFT
Full fine-tuning lets every parameter move. PEFT (parameter-efficient fine-tuning) freezes the pretrained model and trains a tiny adapter — cheaper, safer, and how almost everyone fine-tunes in 2026.
Explain like I'm new to AI
A 7B–70B model has billions of knobs. Turning all of them (full fine-tune) needs a cluster, a full copy of the weights to store, and a high risk of wrecking general skill.
PEFT says: leave the original knobs alone. Glue on a small set of new knobs (an adapter). Train those. Ship base + 50–300MB adapter instead of another 14–140GB clone.
Every parameter can move. Huge GPU bill. You ship a whole new copy of the model.
- + Maximum capacity to change
- − Costly, easy to overwrite general skill
Base weights freeze. A tiny add-on learns the task. You ship the base + a small file.
- + Cheap, swap adapters per product
- − Slightly less room to totally rewrite the model
LoRA is the PEFT method you will actually meet. Prefix-tuning and older adapter blocks exist; LoRA won the ecosystem (vLLM, Ollama, TGI all load adapters).
Mental model
Renovating a house. Full FT is knocking down every wall. PEFT is adding a well-fitted extension. Same foundation; you can unscrew the extension and put a different one on for another product.
How it works
- Full FT: optimizer states for all params (Adam ≈ 8–16 bytes extra per parameter). Multi-GPU. Used by labs, or tiny models.
- PEFT: freeze
W. Train adapter θ. At inference,y = f(W, θ, x). Merge LoRA into W for a single file if you don't need to swap, or keep adapters separate for many tenants.
Multi-adapter: one base, many LoRAs (legal tone vs casual). Load per request. Don't train one mega-adapter for unrelated jobs.
Real-world example
A company has a 8B open model. Support, coding, and HR each get a LoRA. One GPU image. Swap adapter by product. Full FT would be three full models and three times the drama.
Technical explanation
PEFT reduces trainable parameters by orders of magnitude, which cuts VRAM (gradients + optimizer) more than it cuts frozen weight storage. QLoRA then shrinks the frozen storage too.
Trade-off: if you truly need to rewrite the model (new language, extreme domain), full FT or continued pretraining of more layers can beat a rank-8 LoRA. For product SFT/DPO, PEFT is the default until eval says otherwise.
Merging is lossy in the sense you can't unmerge as cleanly if you quantize after — plan the serving stack.
Common mistakes
Full-fine-tuning a 70B on a weekend because "more weights = better." You overfit, forget, and pay the cluster. Try LoRA first.
- One adapter for ten unrelated tasks.
- Forgetting to freeze eval of the base behavior.
When to use it
- PEFT for almost all product fine-tunes. Full FT when you have lab-scale data/compute and PEFT saturates.
When NOT to use it
- Don't full-FT to add a FAQ. Don't PEFT as a substitute for RAG.
Alternatives
- Prompting; distillation into a smaller model; vendor fine-tune APIs (still PEFT under the hood).
Quick quiz
Related concepts
- LoRA and QLoRA — LoRA trains two small matrices instead of the frozen weight W; QLoRA also stores W in 4-bit so a 7B+ model fine-tunes on one GPU.
- What is Quantization? — Quantization stores weights with fewer bits so models fit in memory — 4-bit is common for serving and for QLoRA training, with a small quality trade.
Last reviewed: 2026-09-04 · Written by ByHeart AI · Reviewed by ByHeart AI