One reason the open-source AI community moves so fast is that customizing big models stopped being expensive. The method that did most of that work is LoRA.
The problem
Fully fine-tuning a large model means updating billions of parameters — which needs a lot of GPU memory, produces a full-size copy of the model for every task, and prices out most teams. That was the barrier LoRA removed.
The idea
LoRA (Low-Rank Adaptation) freezes the original model entirely and inserts small trainable matrices — the “low-rank” adapters — alongside it. Only those tiny additions are trained. Because they represent the change compactly, the number of trainable parameters drops by orders of magnitude, and so do the memory and storage costs.
Why it works so well in practice
For a wide range of tasks, a LoRA-tuned model performs close to a fully fine-tuned one. And because each adapter is small and self-contained, you can keep one base model and swap in different adapters for different jobs — a huge practical convenience. The approach also enabled quantized variants (like QLoRA) that fine-tune on even less hardware.
Who should care
Anyone who wants to specialize an open-weight model without a data-center budget. LoRA is the reason “fine-tune your own model” went from a corporate capability to something a hobbyist can do on a single GPU.