Multimodal (LLaVA) support for Android/Mobile?

[Helldez]

Hi everyone,

I was wondering if llama.cpp currently supports multimodal models (like LLaVA) on Android or other mobile devices. I know that llama.cpp has added support for LLaVA, but has anyone successfully run it on mobile?

Some questions I have:

1. Is there official support or any experimental builds for multimodal inference on Android?
2. What are the main challenges in running LLaVA on mobile? (Performance, quantization, etc.)
3. Are there any optimized versions or alternative solutions to make it work efficiently on mobile CPUs/NPUs?

If anyone has tested it or has insights, I’d love to hear your thoughts!

Thanks in advance!