Imagine an AI that can not only understand images and text but also reason across multiple languages. That's the ambitious goal driving the development of large multimodal models (LMMs). But how do we truly test their abilities, especially when it comes to complex, multilingual scenarios? Researchers have introduced a new benchmark called M4U, designed to push these models to their limits. M4U presents a unique challenge: thousands of multiple-choice questions spanning 64 scientific disciplines in Chinese, English, and German. What makes it so tough? The questions require expert-level knowledge and intricate reasoning, often involving diagrams, technical blueprints, and even medical images. It's not enough to simply recognize objects; the AI must understand the relationships between them and draw logical conclusions. The results are surprising. Even the most advanced models, like GPT-4, struggle to achieve high accuracy on M4U. This highlights a significant hurdle in AI development: bridging the gap between language understanding and true, cross-lingual reasoning. The benchmark also reveals a language bias in some models, performing better in English than in Chinese or German. This suggests that training data plays a crucial role, and more diverse datasets are needed to create truly multilingual AI. M4U is more than just a test; it's a roadmap for future research. By identifying the weaknesses of current models, it guides developers towards creating more robust and versatile AI systems. The quest for a multilingual, reasoning AI is far from over, but M4U provides a crucial stepping stone towards that goal.