Model Adapter

Large AI systems, such as those behind chatbots, consist of many billions of numerical values. These values determine how the system responds to an input. Resetting them costs weeks of computing time and millions of euros. A model adapter is the cheap alternative to this. The large system is left completely unchanged, and a small add-on piece, only a few percent of its size, is attached to it. This add-on piece is trained and steers the system’s responses in a desired direction, for example toward legal jargon or a particular coding style.

Why companies don’t need to afford their own model

Without adapters, every adaptation would result in a complete copy of the model. A modern language model quickly takes up several hundred gigabytes of storage. Anyone wanting to serve ten specialist fields would need ten such copies. An adapter, by contrast, is often only a few hundred megabytes in size. You store the base model once and, alongside it, ten small files.

This changes who can even take part. A mid-sized company cannot build a model from scratch. Training an adapter on its own documents, however, is feasible with a single powerful graphics card. A project meant for large corporations thus becomes something a small team can implement in a few days.

There’s also a practical advantage: adapters can be swapped out like attachments on a drill. If an adapter performs poorly, you discard it and keep the base model. You can even run several adapters simultaneously on one server and decide per request which one to use.

What happens inside the model when attaching one

An AI model processes text through many successive layers. Each layer computes using large tables of numbers, known as weights. In the adapter approach, these weights are frozen, that is, fixed and no longer changed. Instead, small additional computational steps are inserted into some layers. Only their values are adjusted during training.

The best-known method is called LoRA, short for Low-Rank Adaptation. The idea behind it: the necessary change to a huge table of numbers can usually be approximated by two very narrow tables. Instead of a million values, you then train perhaps twenty thousand. The result comes astonishingly close to a full retraining, even though less than one percent of the values are altered.

It’s important to distinguish this from two similar approaches. An adapter permanently changes the model’s behavior and is trained for that purpose. A prompt, that is, a cleverly worded instruction in a chat, only takes effect for a single conversation. And a retrieval system that supplies the model with matching documents gives it new knowledge but doesn’t change its style. For specialist vocabulary and tone, the adapter is the right tool; for up-to-date facts, it generally is not.

Adapters in apps, image generators, and cloud services

Adapters are most visible in image generators. In communities around Stable Diffusion, tens of thousands of LoRA files are shared. Each one teaches the model a particular drawing style, character, or image mood. Users download several of them and combine them in a single image request. The files are often smaller than a music album.

With language models, the same thing happens in the background. Major providers offer corporate clients the ability to adapt a model to their own data — technically, this is almost always an adapter, not a new model. The technique is also used on smartphones: a base model resides in the device’s storage, and for tasks such as text summarization or email replies, matching adapters are loaded in addition.

In business news, the term usually comes up in connection with costs. When a report states that a start-up built a specialist model for a few tens of thousands of euros, what’s usually behind it is an adapter on a freely available base model. A common misconception is to mistake this for a standalone model. If the base model is discontinued or the provider changes its license, the adapter alone is worthless.