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
A LoRA adapter is the small file that results from fine-tuning a model with Low-Rank Adaptation. Instead of updating a large weight matrix W directly, LoRA freezes the original pretrained weights and learns a low-rank correction expressed as the product of two much smaller matrices, B and A, so the effective weight becomes W + BA. Because A and B have an inner dimension r (the rank) that is tiny compared to the full matrix size, the number of trainable parameters drops by orders of magnitude, and only those A/B matrices are saved. That is why an adapter is often just a few megabytes to a few hundred megabytes while the base model is many gigabytes. At inference you load the frozen base once and apply the adapter's delta on top; you can keep several adapters for different tasks and swap or hot-load them against the same base, or merge one permanently into the weights. Concretely, a team might ship one 7B base model plus separate small adapters for a support-bot tone, a code style, and a domain vocabulary. When building software with AI, this matters for cost and logistics: training is cheaper and faster because gradients flow through far fewer parameters, storage and distribution are trivial, and multi-tenant serving can share one base across many customers' adapters. The tradeoff is capacity — very large behavioural changes may need a higher rank or full fine-tuning.