The model is a model. It's part of a compression algorithm. The compressed data would be the prompt + choice of which predicted tokens to accept (e.g. when not

The model is a model. It's part of a compression algorithm. The compressed data would be the prompt + choice of which predicted tokens to accept (e.g. when not always choosing the most likely next token). The end-user is supplying the prompt and the choice function is randomized/not being used to store data, thus the end user is providing the compressed data.