Main features are: - Except for the QP solver, all parts are written in plain Matlab. This guarantees for easy modification. Special kinds of kernels that require much computation (such as the Fisher kernel, which is based on a model of the data) can easily be incorporated. - Extension to multi-class problems via error corre cting output codes is included. - Unless many other SVM toolboxes, this one can handle SVMs with 1norm or 2norm of the slack variables. - For both cases, a decomposition algorithm is implemented for the training routine, together with efficient working set selection strategies. The training algorithm uses many of the ideas proposed by Thorsten Joachims for his SVMlight. It thus should exhibit a scaling behaviour that is comparable to SVMlight. cting output codes is included. - Unless many other SVM toolboxes, this one can handle SVMs with 1norm or 2norm of the slack variables. - For both cases, a decomposition algorithm is implemented for the training routine, together with efficient working set selection strategies. The training algorithm uses many of the ideas proposed by Thorsten Joachims for his SVMlight. It thus should exhibit a scaling behaviour that is comparable to SVMlight.