In accelerated MRI reconstruction, recovering anatomy from under-sampled, noisy data is challenging. While deep learning has shown promise, most models rely on convolutions that struggle with long-range dependencies. We introduce VTransFormer, a novel deep unrolled foundation model that integrates Vision Transformers with Channel Attention Blocks. It employs a multi-resolution feature fusion strategy, combining high-resolution spatial detail with low-resolution global context, and dynamically selects between global and local extractors based on adaptive quality evaluation. We comprehensively evaluated VTransFormer on the CMRxRecon Challenge 2025, including multi-center deployment across multiple new centers, multi-disease applications across various cardiovascular conditions, 5T high-field imaging, and pediatric imaging scenarios. Extensive experiments demonstrate significant improvements in SSIM, PSNR, and NMSE metrics across all challenging clinical deployment scenarios, establishing VTransFormer as an effective foundation model for diverse cardiac MRI reconstruction applications.