Miniature unmanned helicopter (MUH) is a controlled member which is very complicated, due to their some characteristics such as highly nonlinear, close coupled, time-variation, open-loop unstable etc. The traditional method of identification is a whole model method. Although those can solve some hard problem, the time-variation is not treated well. The paper introduces a method of model building for miniature unmanned helicopter (MUH), based on local least square support vector machine. Namely the nearest samples to the predicted sample are selected online, and model building is finished by those samples with prediction. The feature of this method is that using the idea of local model building updates the model online, and the global model building brings the low ability of model generalization. In the last, compared with the traditional method of least square support vector machine in the experiment, the results show the algorithm is more effective.