Maybe lots of the data is embedding values or tokenizer stuff, where a single prompt uses a fraction of those values. And then the rest of the model is quite small.