A column's storage cost depends on its values and cardinality, not just its visible width in a table. A unique order identifier needs a larger dictionary than a short Channel column with repeated values. Remove unused columns and rows before loading when the report doesn't need them.
Don't remove a relationship key or a field used by a measure merely because it is hidden. Hidden fields are still loaded. Reducing granularity can save more work than cosmetic changes, but it changes which questions the model can answer. This simulator's Model size pane estimates column storage from distinct values and row counts; it doesn't inspect a compressed Desktop file. Use the explanation and a comparable analyzer capture to reason about the effect rather than quoting its estimate as a production measurement.