LoRA is a technique for fine-tuning large language models efficiently. Instead of retraining all of the model's parameters, LoRA adds small, trainable matrices to specific layers and trains only those — leaving the rest of the model unchanged.
This produces a fine-tuned model that is significantly cheaper to train and easier to deploy than a fully retrained version, while still achieving meaningful domain adaptation. For organizations that want to customize foundation models without the full cost and infrastructure overhead of complete fine-tuning, LoRA is often the right balance of performance and practicality.
