pre-trained model would mean training a language model to predict text, then starting from there and training it to add numbers.

training from scratch would be initializing a neural network, and training it to add numbers directly.