Cardiovascular diseases (CVDs) are the leading cause of death worldwide, and while Cardiac Magnetic Resonance (CMR) is the gold standard for diagnosis, its manual interpretation is labor-intensive and suffers from poor reproducibility—creating a critical need for robust automated analysis. Existing challenges, such as ACDC or LAScarQS, typically focus on single sequences, single views, or isolated tasks like cine segmentation or scar quantification, resulting in models with limited clinical generalizability. To address this gap, we introduce the CMR-Multi Benchmark Challenge: the first large-scale, multi-center dataset that unifies 4D cine (3D+t), late gadolinium enhancement (LGE), and three standard views (SAX, 2CH, 4CH), with expert annotations for ventricles, myocardium, atria, and scar. By integrating temporal dynamics and enabling multi-task evaluation—including segmentation, scar burden quantification, and wall motion abnormality analysis—this challenge fosters the development of cross-domain, generalizable AI models capable of handling real-world variations in scanners, protocols, and patient populations. Aligned with the frontiers of computer-aided diagnosis, CMR-Multi not only tackles key industry challenges like cross-institutional deployment and device heterogeneity but also provides foundational data for cardiac digital twins and medical foundation models, offering significant clinical translation potential and research impact.