Understanding the microstructure of brain tumours without invasive methods remains a major challenge in neuro-oncology. The VERDICT MRI technique provides biologically meaningful metrics, such as cellular and vascular fractions, that help distinguish tumour grades and align closely with histological findings [1,2]. Yet, traditional non-linear fitting approaches are both computationally heavy and prone to errors, which restricts their use in clinical practice. Deep learning presents a promising solution by enabling faster and more reliable diffusion analysis [3]. Still, there is limited evidence on which specific neural network designs are best suited for accurate VERDICT parameter mapping. We present the first head-to-head benchmark of eight neural network families for predicting VERDICT parameters: multilayer perceptron (MLP), residual MLP, Long short-term memory (LSTM)/Recurrent Neural Network (RNN), Transformer, 1D-Convolutional Neural Networks (CNN), variational autoencoder (VAE), Mixture of Experts (MoE), and TabNet. All models were trained and evaluated under a unified protocol with standardized preprocessing, matched optimization settings, and common metrics (coefficient of determination R2, RMSE), supplemented with bootstrap-based uncertainty and pairwise significance testing. Across targets, simple feedforward baselines performed competitively with more complex sequence and attention-based models, indicating that architectural complexity does not uniformly translate into superior accuracy for VERDICT regression on tabular features. Compared to traditional fitting, learned predictors enable fast inference and streamlined deployment, suggesting a practical path toward near-real-time VERDICT mapping. By establishing performance baselines and a reproducible evaluation protocol, this benchmark provides actionable guidance for model selection and lays the groundwork for clinically viable, learning-based microstructure imaging in neuro-oncology.