what-is-lora

---

title: "LoRA (Low-Rank Adaptation): Definition & Uses 2026" description: "LoRA fine-tuning lets you adapt large AI models with a fraction of the parameters. Learn how Low-Rank Adaptation works and why it matters in 2026." slug: "what-is-lora" date: "2026-04-06" updated: "2026-04-06" author: "NovaReviewHub Editorial Team" status: "published" targetKeyword: "LoRA fine-tuning AI definition" secondaryKeywords:

• "what is LoRA in AI"
• "Low-Rank Adaptation explained"
• "LoRA vs full fine-tuning"
• "how to use LoRA for model training"
• "LoRA adapters explained" canonicalUrl: "https://novareviewhub.com/glossary/what-is-lora" ogTitle: "LoRA Fine-Tuning Explained: The AI Technique Saving Time & Money" ogDescription: "LoRA lets you fine-tune huge AI models on a single GPU. Here's how Low-Rank Adaptation works and when to use it." ogImage: "/images/glossary/what-is-lora-og.jpg" ogType: "article" twitterCard: "summary_large_image" category: "glossary" tags: ["LoRA", "Fine-Tuning", "AI Training", "Machine Learning", "LLM Optimization", "Parameter-Efficient Training"] noIndex: false noFollow: false schemaType: "DefinedTerm" term: "LoRA (Low-Rank Adaptation)" definition: "LoRA is a parameter-efficient fine-tuning technique that freezes a pre-trained model's weights and injects small, trainable rank-decomposition matrices, allowing you to adapt large AI models at a fraction of the computational cost." relatedTerms: ["Fine-Tuning", "QLoRA", "PEFT", "Adapters", "Transfer Learning"]

---

LoRA (Low-Rank Adaptation): Definition & Uses 2026

Training a large language model from scratch costs millions. Fine-tuning one used to cost nearly as much — until LoRA changed the math. LoRA (Low-Rank Adaptation) is a technique that lets you customize powerful AI models using a tiny fraction of the parameters, cutting GPU costs by 90% or more while preserving performance.

Whether you're adapting a model for legal document analysis, creative writing, or medical Q&A, LoRA makes it practical to fine-tune on a single consumer GPU. By the end of this glossary entry, you'll understand exactly how LoRA works, when to use it, and how it compares to alternatives like full fine-tuning and QLoRA.

What is LoRA?

LoRA stands for Low-Rank Adaptation — a parameter-efficient fine-tuning (PEFT) method introduced by researchers at Microsoft in 2021. The core idea is surprisingly elegant: instead of updating all the billions of weights in a pre-trained model, LoRA freezes them and injects small, trainable "adapter" matrices into each transformer layer.

Here's the intuition. Imagine a model's weight matrix is a massive spreadsheet with billions of cells. Full fine-tuning changes every cell. LoRA's insight is that the changes you need — the delta — live in a much lower-dimensional space. So LoRA decomposes that delta into two small matrices whose product approximates the full update, but with dramatically fewer parameters.

For a model like LLaMA 3 (70 billion parameters), full fine-tuning requires updating all 70B weights. With LoRA, you might train only 20–200 million parameters — less than 0.3% of the total — and still achieve comparable results on most downstream tasks.