Brain tumors remain a major global health concern, needing accurate diagnostics for effective treatment. This study investigates brain tumor classification on MRI images using EfficientNetV2 variants (B0, B3, and M), with a novel focus on optimizing trainable layer configurations to enhance model adaptability and performance. A publicly available dataset of MRI images with four categories (glioma, meningioma, no tumor, and pituitary) served as the basis for the model training and evaluation. Transfer learning was applied to reduce training time, while data augmentation prevented overfitting and improved performance. Experiments demonstrated that EfficientNetV2B3 achieved the best trade-off between model performance with 99.39% accuracy and computational efficiency, showing strong differentiation between tumor classes with minimal confusion. The model also reached 99.36% F1-Score, making it suitable for balanced approach environments. EfficientNetV2B0 showed faster training time and inference time with slightly lower performance, highlighting its potential for resource-limited scenarios. EfficientNetV2M, the largest model did not outperform the other smaller models when trained on the relatively small dataset. These findings underscore the importance of aligning model complexity with dataset size. By emphasizing both accuracy and feasibility, this research offers to facilitate more reliable and accessible brain tumor diagnosis, improving patient outcomes in diverse healthcare settings.