SmolVLM Support

[cfelicio]

Is your feature request related to a problem? Please describe.

No, but rather a limitation. The current models available on the gallery are great, but require capable hardware for doing vision.

Describe the solution you'd like

Recently, llama.cpp added support for vision (https://news.ycombinator.com/item?id=43943047), and now it's possible to run much smaller models that are good enough for things like surveillance video analysis in realtime (e.g. using SmolVLM)

Describe alternatives you've considered

I'm currently using Gemma3 as an alternative, but it's very hardware intensive.

Additional context

N/A

[mudler]

Update to the new upstream changes is in being worked out in #5379 - that should also bring support for SmolVLM