Multimodal models have a third way: no tokenization, no dual systems

Multimodal models have a third way that does not rely on tokenization or dual-system architectures. Mainstream solutions exist with a gap in the structure of linguistic and visual information, and current models unify these two distributions through innovative methods.