LoRA

LoRA trains a small number of additional weights that sit alongside a base model's original, frozen weights, rather than updating every parameter in the model. That makes it far cheaper and faster to produce than full fine-tuning, and because the base model itself never changes, several LoRA adapters can be trained for different tasks and swapped in over the same underlying model.

A regulated team that wants a model adapted to its own tone, terminology or a specific classification task, without the cost of training and separately hosting a whole model per use case, gets there with a LoRA adapter instead.

Training one adapter for a firm's client-facing tone and a second, separate adapter for internal ticket triage, both served from the same base model rather than two full models.

In Dynamiq, fine-tune LoRA adapters and serve them dynamically over a shared base model, with no separate deployment required per adapter. A request is routed to a specific adapter by name, on the same OpenAI-compatible endpoint the base model itself serves.