MultiNet is a comprehensive benchmark for evaluating multimodal generalization in next generation models across vision, language, and action, revealing critical insights into model capabilities and failure modes at distribution boundaries.