Growing popularity of deep learning (DL) models for classifying motor imagery encephalography signals has driven the development of highly accurate but computationally intensive models. These models pose challenges in deployment on resource-constrained platforms like Internet of Things (IoT) devices due to their high computational requirements. To address this challenge, we propose EEGLite, a lightweight model that combines a minimal convolutional neural network (CNN) with a self-attention module. This CNN architecture reduces trainable parameters and FLOPs, while the self-attention mechanism refines extracted features to preserve accuracy with negligible additional complexity. As a result, EEGLite achieves low computational and energy costs while maintaining high classification performance. EEGLite was evaluated on three benchmark datasets under subject-independent and subject-dependent settings. The results show that EEGLite consistently outperforms state-of-the-art models, including EEGNet, ShallowConvNet, DeepConvNet, EEG-TCNet, EEG Conformer, TCANet and TCNet-Fusion. Notably, on the BCI Competition IV-2b dataset, EEGLite achieves the highest accuracy of 86.23% with only 1818 trainable parameters, while requiring just 1.53 MFLOPs and consuming 3.15 × 105 pJ of energy. Edge deployment was validated on Raspberry Pi 4, achieving 2 ms inference latency, whereas IoT feasibility was validated on the Nordic nRF5340 DK (Cortex-M33, 128 MHz), achieving 27 ms inference latency. RAM and Flash usage were within hardware limits on both platforms. This compact and energy-efficient model facilitates the deployment of brain-computer interface systems in IoT-based healthcare applications, including wearable EEG headsets for home-appliance control and neuro-prosthetic systems, where battery life is a critical constraint.