LoRA Lets Developers Fine-Tune Billion-Parameter AI Models on a Single GPU

LoRA, or Low-Rank Adaptation, is a technique that fine-tunes large AI models by training only a small fraction — often under 1% — of their parameters. Instead of updating every weight in a model, LoRA learns two small matrices whose product approximates the weight changes needed for a new task, leaving the original model frozen. This drastically reduces memory requirements, making it possible to fine-tune models with 7 billion or more parameters on a single consumer GPU rather than an expensive cluster. Each trained adapter is a standalone file of just a few megabytes, allowing a single shared base model to serve many specialized variants by hot-swapping adapters at runtime. Alternatively, the adapter can be permanently merged back into the base model, eliminating any added computational overhead during inference.