Adapters are an advanced method for making pre-trained AI models adaptable to new tasks without complete retraining. These modules save time, money, and resources by efficiently repurposing existing models for different tasks in areas like natural language processing, computer vision, and robotics.
Full fine-tuning rewrites every weight in a model. For anything at modern scale that means copying billions of parameters, paying for the compute to update all of them, and storing a complete new model for every task you care about. Adapters sidestep that entirely: the original weights stay frozen, and a small set of new ones learns the task instead.
How adapters work
An adapter inserts small trainable modules into a frozen model — typically a handful of layers with far fewer parameters than the network around them. During training, gradients only flow into those modules. The base model is untouched.
The dominant variant is LoRA (Low-Rank Adaptation), introduced by Microsoft researchers in a 2021 paper. LoRA represents the weight update as the product of two much smaller matrices, which means you can often train well under 1% of the model’s parameters and still match full fine-tuning on the target task. Hugging Face’s PEFT library packages LoRA alongside related methods like prefix tuning and IA³.
Why adapters matter
Cost. Training a LoRA on a mid-sized open model is a single-GPU job measured in hours, not a cluster job measured in weeks. For most teams this is the difference between fine-tuning being an option and not.
Storage. A trained adapter is often a few megabytes against a base model’s tens of gigabytes. You can keep dozens of task-specific adapters — one for support tickets, one for product copy, one for code review — and swap them against a single base model at inference time.
Reversibility. Because the base weights never change, an adapter that misbehaves is deleted, not recovered from. Full fine-tuning gives you no such undo.
The trade-off: adapters are good at teaching a model a style, a format, or a domain vocabulary. They are much weaker at teaching it facts it never saw during pre-training. If your problem is that the model does not know your current product catalogue, the answer is usually retrieval-augmented generation, not an adapter.
Related terms: Fine-tuning, Parameter-Efficient Fine-Tuning (PEFT), Foundation Model, Transformer.
