Early models worked with a single kind of input. Multimodal models take several at once, such as describing a photo, answering a question about a chart, or responding to spoken words, by representing each kind of input in a shared form the model can reason over.
This lets a user point at a picture and ask about it, or mix a diagram with a written question. The difficulty is aligning very different inputs so the model relates what it sees to what it reads, rather than treating them separately.