We introduce a fast and theoretically founded method for learning landmark-based SVMs in a multi-view classification setting which leverages the complementary information of the different views and linearly scales with the size of the dataset. The proposed method – called MVL-SVM – applies a non-linear projection to the dataset through multi-view similarity estimates w.r.t. a small set of randomly selected landmarks, before learning a linear SVM in this latent space joining all the views. Using the uniform stability framework, we prove that our algorithm is robust to slight changes in the training set leading to a generalization bound depending on the number of views and landmarks. We also show that our method can be easily adapted to a missing-view scenario by only reconstructing the similarities to the landmarks. Empirical results, both in complete and missing view settings, highlight the superior performances of our method, in terms of accuracy and execution time, w.r.t. state of the art techniques.