Low-Rank Adaptation, or LoRA, reduces the number of parameters that must be trained to adapt a model. Instead of updating every weight, it adds trainable low-rank matrices to selected layers while leaving the base weights mostly unchanged. This can reduce training memory and make task-specific adapters easier to store.
In practice a team can keep one base model and swap small adapter files for different tasks, such as a legal-style adapter and a support-tone adapter. Because only the small matrices are trained, experiments are cheaper than updating every weight. Track which base model version each adapter was trained against, since adapters do not transfer automatically to a different base.
LoRA still requires suitable training data and evaluation; it does not guarantee better performance. The quality depends on the base model, target task, adapter configuration and data. Related: Model Fine-Tuning, Open-Weight Model.