This paper investigates the minimax problem of a non differentiable saddle functional,and constructs a nonsmooth neural network to search for the saddle point of a saddle functional.Under suitable conditions the convergence and stability of the network are discussed,using Lyapunov theory.