Automatic tagging of images is what I'm doing with it. I'm using a modified Gemma model, which ingests images (and audio, for that matter).