Unsupervised clustering of high-dimensional complex datasets is still challenging in machine learning. This study proposes deep mixtures of variational autoencoder’s (MVAE) model to learn meaningful latent representations for unsupervised cluster analysis. The MVAE model extends the existing deep clustering algorithms by assuming the mixture’s distribution for both the variational posterior and prior components of the VAE framework. We derive an evidence lower bound (ELBO) to approximate the marginal log-likelihood function of our proposed model and formulate a closed-form expression for estimating the clustering assignment probabilities. Our optimization equation explicitly integrates the clustering distribution with the regularization and the reconstruction component. The proposed mixtures VAE model achieves superior clustering performance compared to baseline algorithms on various benchmark datasets. The MVAE model also demonstrates the ability to generate realistic examples using samples from the latent space.