The brain tumor grade classification is one of the prevalent tasks in brain tumor image classification. The existing models have employed transfer learning and are unable to preserve semantic features. Moreover, the results are reported on small datasets with pre-trained models. Thus, there is a need for an optimized model that can exhibit superior performance on larger datasets. We have proposed an efficientnet and coupled convolution network for the grade classification of brain magnetic resonance images. The feature extraction is performed using a pre-trained EfficientNetB0. Then, we have proposed a coupled convolution network for feature enhancement. Finally, enhanced features are classified using a fully connected dense network. We have utilized a global average pooling and dropout layers to avoid model overfitting. We have evaluated the proposed model on the REMBRANDT dataset and have achieved 96.95% accuracy. The proposed model outperforms existing pre-trained models and state-of-the-art models in vital metrics.