Put images and audio before the text
prompting & control tokens
// what it does
In a multimodal message, place image and audio parts before the text instruction. Gemma 4 was trained with media tokens ahead of the question, so text-first prompts measurably weaken visual and audio grounding. Audio inputs are limited to about 30 seconds and to the E2B/E4B/12B sizes, and video is processed as sampled frames.
// gotcha
The ordering is a training convention, not a hard API error — a text-first prompt still runs but returns worse captions, OCR, and answers, which is easy to misdiagnose as a model-quality problem.