Graph-theoretical representations for sets of probability measures (credal networks) generally display high complexity, and approximate inference seems to be a natural solution for large networks. This paper introduces a variational approach to approximate inference in credal networks: we show how to formulate mean field approximations using naive (fully factorized) and structured (tree-like) schemes. We discuss the computational advantages of the variational approach, and present examples that illustrate the mechanics of the proposal. Keywords. Credal networks, variational methods, inferences.