All guides
Models· 3 min

How does LoRA work?

LoRA fine-tunes a big model cheaply by learning small low-rank update matrices instead of touching all weights.

Key idea

Freeze the base model; add tiny trainable rank-r matrices A·B to each layer. Train those; keep the base intact.

Example

Fine-tune a 70B model on your product docs by training ~0.1% of parameters on a single GPU.

What to remember

  • Drastically fewer parameters to train
  • Adapters swap in/out at inference time
  • Basis for QLoRA (quantized) and LoRA merges