The demand for efficient utilization of computational resources on Low Earth Orbit (LEO) has driven the need for a cloud-edge model collaboration framework. This paper proposes a novel framework that leverages reinforcement learning algorithm to compress the original model, thereby reducing the computational overhead on limited computing resources. The original and compressed models are then deployed using containerization techniques. Finally, the collaborative model inference between the cloud and edge is achieved through the utilization of the KubeEdge. The application of this framework holds significant promise on LEO, enabling optimized utilization of computational resources while maintaining reliable and efficient model inference. The use of reinforcement learning-based model compression techniques contributes to reducing the computational complexity. The containerization deployment approach ensures flexibility, portability, and seamless integration with existing infrastructure. This comprehensive framework presents a proposed solution for mitigating the challenges associated with resource constraints and facilitating model collaboration on LEO. The experimental evaluation demonstrates the effectiveness and performance benefits of the proposed framework, paving the way for its practical implementation and deployment in LEO applications.