Timely and accurate flash flood early warning is one of the important techniques for flash flood defense, which is crucial for reducing the impact of flash floods. Currently, the commonly used flash flood warning approach is the rainfall threshold approach, which belongs to the deterministic early warning approach. However, due to the complexity and uncertainty of flash floods, the deterministic approach has some limitations in practical application. Based on the parameter uncertainty of hydrological model, a probabilistic warning model for flash floods is proposed. First, the Monte Carlo method was applied to the Xinanjiang model to analyze the parameter uncertainty, and then the rainfall threshold was calculated for each time span for the selected parameter sets, and the probability of exceedance was determined by combining with the probability density analysis. Taking the flash flood warning in Qingxi River catchment of Sichuan Province as an example, the results show that compared with the deterministic approach, the number of false alarms is significantly reduced from 28 to 16, and the critical success index (PCSI) is improved from 0.41 to 0.54, which indicates that the probabilistic warning approach has a high accuracy and reliability, and it can be better applied for flash flood warning.