Gemma3 models
Do the GGUF quants of Gemma3 models contain vision layers?
Do the GGUF quants of Gemma3 models contain vision layers?
For the few Gemma3 models where the original model contains a vision layer we provide the mmproj file for llama.cpp vision capability in the static quant repository in Q8_p0 and F16 as can for example be seen under https://huggingface.co/mradermacher/amoral-gemma3-12B-vision-GGUF/tree/main
Keep in mind that vision is never inside the GGUF itself but always in the separately provided mmproj files. If you want vision you need to provide llama.cpp with booth the GGUF and mmproj file.
Do all Gemma3 models use the same mmproj files? (for the same size)
Do all Gemma3 models use the same mmproj files? (for the same size)
We always generate new model specific mmproj files when we see that the model we quantize supports vision capability. All the quants of the same model can use any mmproj file provided in the static quant repository of that specific model. Q4_K_M imatrix quants for example support booth F16 and Q8_0 mmproj files. If the author didn't finetune the vision part you might be able to take the mmproj file from the base model instead of redownloading but I see no reason why you would want to do so as it will only cause you issues. There are in fact two Gemma3 architectures Gemma3ForConditionalGeneration with vision and Gemma3ForCausalLM without vison so for models that don't support it is not just the vision layers are missing but the architecture itself is not supporting vision so just adding some other models mmproj will probably not work.