The time step is a critical parameter for the stable operation of numerical weather models. To achieve stable model operation and ensure efficiency, an adaptive time step scheme is designed based on the characteristics of the time integration scheme of the CMA-MESO 3 km model. This scheme relies on the model’s maximum Courant number and proposes two methods: the target aiming method and the top trimming method, depending on the adjustment approach. (1) Target aiming method: after adjusting the model’s time step, the maximum Courant number in the model approaches a given target value throughout the entire integration process. (2) Top trimming method: after adjusting the model’s time step, the maximum Courant number in the model does not exceed a given target value throughout the entire integration process. In order to implement the adaptive time step scheme in the CMA-MESO 3 km model, time control technology suitable for the CMA-MESO 3 km model is developed, including time control for the integration process, time processing for input/output, and time interpolation handling. Two types of experiments, namely case experiments and batch experiments, are designed to verify the application performance of the scheme in the CMA-MESO 3 km model. The individual case experimental results show that when the target aiming method is employed, the maximum Courant number in the model fluctuates around the target value, enhancing model stability. As the target Courant number decreases, the model becomes more stable, but the total number of integration steps and the total integration time increase. When using the top trimming method, adjustments to reduce the time step are only made when the maximum Courant number in the model exceeds the target value, keeping the maximum Courant number close to the target. When the Courant number is below this target value, if the model’s time step has been previously adjusted, the time integration step is gradually restored to the initial time step; otherwise, the model’s time step remains unchanged. The results of batch experiments indicate that the adaptive time step scheme of both methods can effectively avoid the situation of model integral overflow and significantly improve the stability of model integration. The impact of the scheme on the simulation results of model precipitation and geopotential height fields is relatively small. For a 3 km resolution model, when the target Courant number is set at approximately 1.2, the top trimming method is more suitable for operational use than the target aiming method. This scheme not only ensures the stable operation of the CMA-MESO 3 km model but also allows a larger integral time step, thereby enhancing the operational efficiency of the model. Currently, the adaptive time step is implemented in operational applications.