Earthquakes can result in strong, unforeseen releases of energy on the earth’s surface, destroying life and property. The impact of an earthquake can be studied based on onsite earthquake magnitude. Traditionally, the magnitude estimation regression relationships have been used to predict earthquake magnitude based on several dominating parameters. These dominating parameters are predicted from the early three seconds of the primary waveform. Deep learning techniques have been proven helpful for real-world applications. Therefore, in this study, DeepMag, a single-station deep learning prediction model, has been designed to predict the earthquake magnitude using Long Short-Term Memory (LSTM) followed by a machine learning model such as Light Gradient Boosting Machine (LightGBM). The parameters of the LSTM layers in the proposed model are predicted using a genetic algorithm. The DeepMag model has been fed with the complete three-second recordings of three waveform components corresponding to the primary seismic signal. Additionally, various significant parameters derived from this waveform have been included in the model input data. The dominating parameters include eleven parameters obtained by applying the Butterworth and Chebyshev filter on three seconds of the vertical component of the seismic waveform. A total of 2094 acceleration records obtained from earthquakes recorded by an early warning system installed by Indian Institute of Technology, Roorkee have been used in the present work. The dataset was divided into 1465, 440, and 189 records for training, testing, and validation of the developed deep learning model with 70:21:9% split. The hyperparameter of the model has been obtained using grid search. The study shows the mean absolute error of 0.48 obtained from DeepMag. An ablation study of the model has been conducted by employing various state-of-the-art deep learning model-based prediction techniques. The results demonstrate that the proposed methodology performs better than the recently developed state-of-the-art models. The multi-station model analysis was performed for three earthquakes (5.2, 6.1, and 6.2 Mw). It is seen that the mean error of -0.28, 0.01, and 0.04 Mw in magnitude prediction was obtained for this earthquake using multi-station recorded data. The present study shows that the DeepMag model can predict the magnitude using the first 3 s of single and multi-station recordings, thereby establishing the efficacy of the developed deep learning model.