Ollama has had vision support for Gemma3 since it came out. The implementation is not based on llama.cpp's version.