LoRA, short for Low-Rank Adaptation, is a parameter-efficient way to adapt a large pretrained model to a new task without updating all of its original weights. Instead of retraining the whole model, LoRA learns a much smaller set of added parameters that modify how the base model behaves.
Why LoRA Became Popular
Large models are expensive to fine-tune directly. They require significant memory, compute, and storage, especially if many customized versions are needed. LoRA became popular because it makes adaptation cheaper and more practical while still producing strong task-specific improvements in many settings.
This is particularly helpful when organizations want specialized variants for different domains, styles, or customers. Instead of duplicating huge models repeatedly, they can keep one base model and apply lighter-weight adaptations.
What LoRA Means in Practice
LoRA is not magic. It is one design choice inside the larger space of transfer learning and fine-tuning. It often works well, but its value depends on task fit, evaluation quality, and deployment needs. Some tasks may still require fuller retraining or stronger data preparation.
Even so, LoRA is an important concept because it shows how modern AI engineering often progresses: not only through bigger models, but through smarter ways to adapt and operate them.
Related concepts: Fine-Tuning, Transfer Learning, Large Language Model (LLM), and Machine Learning.