Can someone recommend me a local harness to explore the multimodality of this kind of models? Text in text out is mostly generically solved, but that's about what I know