LoRA (Low-Rank Adaptation) is the efficient-training workhorse: instead of touching billions of base parameters, only a few small added matrices are trained. Fine-tuning cost drops from “cluster” to “one GPU card.”
Why it is everywhere: multiplicity (dozens of LoRAs on one base model for a dozen tasks), speed (hours) and reversibility (drop back to the base). In open-source ecosystems, LoRA means the democratization of customization.
Care points: stacking many LoRAs degrades quality; and for bulk knowledge injection, RAG is still wiser. LoRA is built for behavior and style, not for knowledge bases.
