Modern autonomous driving systems rely on accurate and reliable future motion predictions from deep neural networks to ensure safe and efficient navigation. However, neural networks often produce overconfident predictions, especially under high situational uncertainty, which can result in catastrophic failures in safety-critical applications. To mitigate this risk, it is essential for models to demonstrate that they can effectively handle uncertainty effectively. In this work, we focus a state-of-theart, efficient, query-based occupancy grid prediction model known as ImplicitO [2]. Occupancy grid prediction plays a vital role in many autonomous driving systems by providing a compact cell-wise representation of the surrounding environment. This makes it especially well-suited for the estimation of uncertainties. We investigate various methods to enhance ImplicitO with uncertainty mechanisms. We evaluate the performance of different model configurations, including extensions with standard dropout, concrete dropout (CD), a conditional variational autoencoder (CVAE), and aleatoric heteroscedastic uncertainty prediction. We analyze our models both quantitatively and qualitatively to assess the efficacy of these methods. Our results indicate that using concrete dropout improves the predictive performance and robustness even if larger datasets are available. The CVAE models struggle with accurately predicting uncertainty, especially for longer prediction horizons, while heteroscedastic extensions, in particular a model that additionally predicts a misclassification (label-flip) probability, show potential in capturing predictive uncertainty, reflected by higher predictive entropy and improved calibration for future time steps. Heteroscedastic models are compatible with concrete dropout and show promising results when combined.