This paper introduces a variational formulation for Maximum Mean Discrepancy, a generative modeling framework based on RKHS techniques. The formulation is given in terms of a gradient flow in 2-Wasserstein space. While the gradient flow viewpoint is not particularly new in the context of generative modeling, the scope and the quality of the results (convergence toward global optimum; regularization by noise injection; closed-form implementation of the updates) are sufficient for a poster presentation at NeurIPS.