I think they mean whatever’s handling the model. A program into which you feed this inherently restricted format, so it takes advantage of those limitations, in order to run more efficiently.

Like if every number’s magnitude is 1 or 0, you don’t need to do floating-point multiplication.