Which mmrpoj file for Beepo 32b? (or which one for Qwen 32b?)

[illtellyoulater]

A couple of questions:

• README recommends Beepo-22b as the largest and most powerful LLM. Is there an mmproj that can be used with it? Or are there any plans to release one?

• yi-34b is the biggest LLM supporting vision through mmproj. However Qwen 32b is the highest scoring 32b LLM (and the only one capable of even surpassing a few 72b models). So is there an mmproj that could be used with it? Or do you plan releasing one?

Thanks!

[LostRuins]

Unfortunately while there is Pixtral, nobody has bothered to add llama.cpp support for it yet. So for now your best options will be Qwen2VL or MiniCPM.

[illtellyoulater]

I’m sorry, but I’m not sure which part of your answer applies to each of my questions.
Could you clarify, please?
Thanks!

EDIT - making my way through it...

• Beepo-22b and mmproj: as of now, there’s no mmproj available or compatible with Beepo-22b.
• Qwen-32b and mmproj: while Pixtral is a multi-modal framework, it currently lacks llama.cpp support. This makes using Qwen-32b with mmproj infeasible at the moment.

So, the best available paths for now are:

• MiniCPM (SOTA vision model paired with a small LLM)
• Qwen2VL (good vision model paired with a significantly larger LLM)
• yi-34b (the only ~30b multi-modal solution for now; not sure how good the vision is for this one, but perhaps the best compromise, overall, for now)