[upboundspiral]

I've been using Qwen 3.5 35B-A3B with images as input so I suspect you perhaps didn't include the vision part of the model during testing (I use llama.cpp and I learned I needed to include the separate mmproj part).