Term
LoRA adapter
The small file produced by LoRA fine-tuning. Contains only the rank-decomposed updates to a few attention layers, typically 10-200 MB. Can be loaded onto the base model at inference for a customised checkpoint without storing the whole model.
Background
LoRA adapters are how the open-weights image-gen and small-LLM ecosystems share custom styles cheaply. Civitai hosts tens of thousands of LoRAs for Stable Diffusion. For LLMs, adapters are less common (frontier models aren't open) but show up heavily in research and on-prem deployments of Llama and Qwen.