Technique
QLoRA
LoRA on a 4-bit quantised base model, enabling 65B fine-tunes on a single 24GB GPU.
Definition
QLoRA combines NF4 quantisation of the frozen base with LoRA adapters. Memory drops 4–8× versus FP16 training, putting fine-tuning of huge models within reach of researchers and indie developers. It was a major democratising step for open-weights work.
Common use cases
- Single-GPU fine-tuning
- Research
- Indie deployments