It's possible they have a software layer that does that. But I was assuming they don't, because the open source multimodal models don't.

No it’s not, it’s a multimodal transformer model.