A parallel elastic actuator (PEA) leg is proposed to improve the motion performance of quadruped robots. By imitating the tendon-driven mechanism of quadruped animals, a tension spring is connected in parallel between the calf link and the thigh link to achieve parallel elastic effect at the knee joint. The PEA leg exhibits small size, low mass, and low inertia characteristics. A dynamics model of PEA knee joint is established. The proposed model is used for a multi-objective optimization of spring stiffness in considering peak torque, peak power, and energy efficiency in a predefined trajectory tracking task. The most optimal spring stiffness is determined to be 5510N/m. The performances of PEA joint and rigid actuator (RA) joint are compared through simulation and prototype experiments. A prismatic joint with stiffness is employed to simulate the effect of PEA tension spring in the Gazebo robot simulation environment. The output torque, power, and mechanical energy consumption of knee motor are obtained through physics engine simulation. A PEA knee joint experimental platform is constructed, and the prototype experiments are conducted to obtain the electrical parameters of the knee joint motor, such as input current, input power, and electrical energy consumption. The experimental results demonstrate that, compared to RA joint, the PEA joint achieves significant improvements. Specifically, within the trajectory period from 0.5s to 2.0s, the PEA joint reduces peak torque by 40% to 79%, peak power by 52% to 89%, and mechanical energy consumption by 40% to 89% in terms of motor output. Regarding motor input, the PEA joint reduces peak current by 46% to 77%, peak power by 43% to 76%, and electrical energy consumption by 62% to 73%.