Google calls Gemini 3 "the best model in the world for multimodal understanding," and it's widely rolling out in preview.