Model adaptation is the process of adjusting a pretrained model’s parameters or behaviour to perform well on a new task, domain, or dataset distinct from the one it was originally trained on. It spans a spectrum from full fine-tuning of all parameters to parameter-efficient methods such as LoRA, which inject small trainable matrices into frozen layers to adapt behaviour at a fraction of the memory and compute cost. Effective model adaptation must balance plasticity, the ability to acquire new task-specific behaviour, against catastrophic forgetting of prior capability. It is the practical mechanism through which large pretrained foundation models are specialised for downstream applications.