QLoRA

Also known as: Quantized LoRA

QLoRA combines 4-bit (or similar) quantization of the base model with LoRA adapters trained in higher precision. The frozen base stays memory-cheap; gradients flow through the adapters.

It is popular for fine-tuning 7B–70B models on a single consumer or mid-tier GPU. Quality depends on quantization scheme, LoRA capacity, and data — not magic by itself.

← Back to AI Glossary