How does LoRA work?

Rather than updating all of a model's billions of weights, LoRA freezes the base model and trains small low-rank decomposition matrices (typically containing 0.1%-1% of the parameters) that are inserted into attention layers. The result is a small adapter file (often just a few megabytes) that can be applied on top of the base model at inference time, dramatically reducing storage, training cost, and serving complexity. The platform strengthens enterprise re