Comment by helloplanets

And most LLMs have been multimodal for years at this point.

Even if the input is in plain English, the model never sees any words, tokens or glyphs to begin with. It's vectors all the way down.