Plant disease detection is one of the critical applications of modern agriculture as it holds the power to threaten food security of millions, yet traditional methods by inspecting the plants physically are labor-intensive and subjective. The paper proposes a more optimized lightweight deep learning model based on MnasNet to overcome the limitations of computationally constrained environments with excellent predictive performance. With merely 3.18 M parameters, we achieve a classification accuracy of 96.14%, thus validating the feasibility of achieving high performance and robustness with significantly reduced complexity. Key architectural innovations include the integration of Squeeze-and-Excitation modules for enhanced feature recalibration and optimized transfer learning techniques that ensure robust and efficient feature extraction. The ultra-light weight design sets this model apart from state-of-the-art alternatives which utilize at minimum tens of million’s parameters at the lower bound. This study showcases the potential of resource efficient architectures to revolutionize plant disease diagnosis by enabling real time, mobile-ready applications.