# What is LoRA fine-tuning?

LoRA (low-rank adaptation) fine-tunes a model by training small adapter matrices on top of frozen weights instead of updating every parameter. It captures most of the task gains of full fine-tuning at a fraction of the GPU cost and memory, which is why it is the production default for fine-tuning.

The base model stays untouched; the adapter is a separate artifact, usually well under 1% of the model’s size. That makes runs cheap to repeat, easy to roll back, and simple to stack: one base can carry a different adapter per workload.

Cost is the practical argument. At around 2,200 training tokens per second per H100-class GPU, a three-epoch LoRA run over a few million tokens finishes in hours, not days. Put your dataset size through the fine-tuning cost calculator for the exact bracket.

## See also

- [Fine-tuning cost calculator](https://allocate.network/tools/fine-tuning-cost-calculator)

## Related terms

- [Fine-tuning](https://allocate.network/glossary/fine-tuning.md)
- [Model weights](https://allocate.network/glossary/model-weights.md)
- [QLoRA](https://allocate.network/glossary/qlora.md)

---

[HTML page](https://allocate.network/glossary/lora) · [Machine-readable catalog](https://allocate.network/catalog.json)