A pre-trained language model whose weights are not updated during inference or deployment, only its outputs are modified.