Ground motion models (GMMs) are traditionally developed from a frequentist approach. The Bayesian framework has received recent attention in developing nonergodic models, measuring uncertainty, or updating the model with additional data. However, no neural networks are developed to date in this framework to predict ground motion parameters or spectra. Hence, the present work develops a probabilistic Bayesian neural network (PBNN) to next‐generation attenuation – West2 and Subduction databases using variational inference with mean‐field assumption. Network inputs are magnitude, rupture distance, hypocentral depth, shear wave velocity, style of faulting, and region flags; outputs are peak ground values and response spectra. Both models have two hidden layers with seven neurons in each hidden layer. The models are verified for potential overfit, and their performance is validated through the parametric study by varying inputs. The output of a deterministic model is a point estimate. Considering probabilistic layers in hidden and output layers enables the model to capture within‐model epistemic uncertainty and aleatory variability. Obtained aleatory standard deviations are consistent with other models. Mean epistemic uncertainty and aleatory variability are in the range 0.07–0.10 and 0.62–0.78 (ln units) for NGA‐West2 and 0.09–0.16 and 0.67–0.95 for NGA‐Sub models, respectively. The correlation coefficients between recorded and overall mean predictions ranged from 0.94 to 0.97 for NGA‐the West2 model and from 0.91 to 0.95 for the NGA‐Sub models. Network performance for out‐of‐training inputs showed increased epistemic deviations with no effect on aleatory deviations. [ABSTRACT FROM AUTHOR]