A new multi classifier was proposed based on support vector machines for a N class classification problem, which comprised N 1 support vector machines in the form of a binary tree. The generalization performance of multi classifiers was discussed, and a new learning algorithm, the BTSVM algorithm, was presented based on high dimension feature spaces. The BTSVM algorithm evaluates example distances by kernel functions, employs the maximization of minimum distances as clustering criteria to obtain two optimal subsets, and generates the optimal classification functions with support vector machine leaning algorithms at each decision node. Theoretical analysis and experimental results show that the BTSVM algorithm is superior to other competitive multi classifiers.