LoRA

in English: Low-Rank Adaptation

professionalwidespread

A resource-efficient fine-tuning method that leaves the large AI model untouched and trains only a small, interchangeable add-on for the new task.

In plain terms

Imagine you have a multi-thousand-page technical book and want to adapt it to your own field. You could rewrite the whole thing, but it is much cheaper not to touch the book and just stick adhesive notes on the relevant pages. LoRA does exactly this with an AI model: the original model remains unchanged, and the notes are small, separately trained add-ons.

This is necessary because a large language model consists of billions of adjustable numbers (parameters), and retraining all of them is expensive, slow, and requires enormous computing resources. With LoRA, only a fraction of these need to be trained, allowing a company or even an individual to shape a model to their own taste, often on a single powerful graphics card.

The add-on (often called a LoRA adapter) is a small file, typically ranging from a few megabytes to a few hundred megabytes, while the model itself can be tens of gigabytes. You can keep multiple adapters for the same base model, for example one for legal texts, one for customer service tone, and swap them like camera lenses.

How it works

The original parameters of the base model are "frozen," meaning none of them change during training.
Two small number tables (matrices) are attached to each selected layer of the model. Their product determines by how much the behavior of that layer should be adjusted.
'Low rank' means these tables are intentionally narrow. The assumption is that changes needed for a new task can be compressed into a few main directions, so the full, massive table is not required.
During training, only these small tables learn from your own examples, often representing less than one percent of the total parameter count.
During use, the adapter is either loaded alongside the model or permanently merged into it. In the latter case, response speed is no slower than with the original model.

An everyday example

On an image generation site, you select a 'watercolor style' or an add-on that knows a specific cartoon character, and your images are immediately produced in that style. This small downloadable file is very often a LoRA: the same base model runs, just with different note sheets.

Who is affected

developerThis is the most common way to customize open models on affordable hardware and maintain multiple task-specific adapters for a base model. → Code

entrepreneur, executiveYou can get a model tuned to your company's terminology and tasks at a fraction of the cost, even if your data never leaves your own machines.

marketerIn image generation, LoRA can teach the model your brand's visual style, a product, or a consistent character. → Marketing

lawyer, citizenFrom a few photos, a LoRA can be created for anyone's face or an artist's style, raising copyright, personality rights, and deepfake concerns. → Law

A common misconception

MythLoRA is a standalone, small AI model that can be run independently.

In realityLoRA is only an add-on: it always requires the base model it was created for and is useless without it; it typically does not work with other models or model versions.

Limits and risks

LoRA mainly teaches style, tone, and task patterns; if the model needs to know fresh or frequently changing facts, a document-search-based solution (RAG) is generally more suitable.
For very complex tasks or those far from the base model's knowledge, it may deliver weaker results than full fine-tuning.
With few or one-sided training examples, the model "overfits": it repeats the examples and performs worse elsewhere.
The origin and content of adapters downloaded from the internet may be unverified; moreover, an adapter can weaken the base model's built-in safeguards, so use them only from trusted sources.

Similar, but not the same

Full fine-tuning – There, every parameter of the model is changed and a whole new copy of the model is created, whereas LoRA trains only a small add-on alongside the untouched model.
RAG (retrieval-augmented generation) – RAG does not train the model but provides external documents at response time, whereas LoRA changes the model's behavior itself.
LoRa (radio technology) – Written with a lowercase "a", LoRa is a long-range, low-power radio data transmission technology for smart devices and sensors (the LoRaWAN network standard is built on this), and has nothing to do with artificial intelligence.

Related concepts

Broader

Fine-tuning, parameter-efficient fine-tuning, Transfer learning

Narrower

QLoRA, DoRA, LoRA-adapter

On the edge

Foundation model, Quantization, RAG, Open-source model, Diffusion model

A brief history

The method was described by Microsoft researchers (Edward Hu and colleagues) in 2021, originally for cheaper customization of large language models. It was soon adopted by users of image-generation models, who began using it to teach styles and characters.

Where things stand

At the end of 2026, LoRA is the default method for customizing open models, for both language and image-generation models. Major developer toolkits support it natively, and several cloud providers' fine-tuning offerings are built on it.

Status: 2026 November

Also known as

LoRA · low-rank adaptation · low-rank adaptation · low-rank fitting · LoRA-adapter · LoRA fine-tuning · LoRA fine-tuning · LoRA-modell

Sources

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.