Evaluating properties of a probabilistic model, such as means, variances and correlations from probability distributions is inherently challenging due to its complexity, here defined as the number of operations and memory allocations required, which often scale exponentially in the number of variables involved. Indeed, just thing about the Ising model and the Potts model, heavily used in the statistical physics context and related fields, for which the evaluation of the partition function Z, necessary for the evaluation of marginals, requires an exponential number of operations, i.e. for N Ising spins the calculation of Z requires 2^N operations. Tensor networks (TNs) offer a novel approach to handle probability distributions, enabling the use of optimization techniques and heuristic strategies to significantly reduce the complexity of the calculation of marginals and normalisation constants of whatever the probabilistic model of interest is.