In this paper, to improve the accuracy of the one-mass and three-mass inverted pendulum models, which have been used for generating real-time walking patterns for biped robots, we propose a novel model based on the three-mass inverted pendulum. The proposed model employs an approximation of moment of inertia of the swing leg to improve the accuracy of the three-mass inverted pendulum in estimating dynamic behavior of the robot. In order to show significance of the proposed model, trajectories for the Center of Mass (CoM) are obtained using the three models, based on a desired ZMP trajectory. The task space trajectories, then, are mapped into the joint space, using inverse kinematics. Having the joint space variables, the actual ZMPs for the three obtained walking patterns are computed and compared. This comparison well shows merit of the proposed model in estimating dynamic behavior of the robot, especially for walking with relatively high speeds. The kinematic and dynamic properties of the models in this paper are based on the humanoid robot SURENA III, which has been designed and fabricated in the Center of Advanced System and Technologies (CAST), university of Tehran.