“vision-based” large language models.

Odd restriction. Why not investigate text-based ones?

Or is “vision-based” a technical term that encompasses models that were trained on text?