QLoRA (Quantized Low-Rank Adaptation) is an efficient fine-tuning method for large language models. It works by quantizing a pretrained model to 4-bit precision and freezing it, then training only the added low-rank adapters (LoRA). This method achieves lossless fine-tuning of 4-bit quantized models for the first time, enabled by three key techniques: the 4-bit NormalFloat (NF4) quantization — an information-theoretically optimal data type for normally distributed weights; double quantization, which further quantizes the quantization constants to save additional memory; and paged optimizers, which prevent memory spikes during training via unified memory paging. The Guanaco model family, trained with QLoRA, achieves strong performance on the Vicuna benchmark — for instance, Guanaco 65B reaches 99.3% of ChatGPT's performance after fine-tuning on a single 48 GB GPU for 24 hours, dramatically lowering the memory barrier for large-model fine-tuning. Key conclusions from the paper include: data quality far outweighs data quantity, and QLoRA can fully recover the performance of 16-bit full-parameter fine-tuning.