Google's Lookout app (accessibility for the blind and visually impaired) was updated ~6 months ago with a multimodal LLM already.

It uses the Flamingo model family: https://deepmind.google/discover/blog/tackling-multiple-task...