Multi-model ensemble (MME) from various global climate models (GCMs) is widely used to quantify hydrological impacts of climatic change from local to global scales. Convolutional neural network (CNN) is a powerful tool for efficiently extracting hidden spatial information from huge datasets, and has been increasingly implemented in geoscientific studies. In this study, a CNN framework is developed to perform MME of CMIP6 monthly precipitation over China particularly at local scale, compared to 32 involved GCMs, quantile mapping (QM), and other prevalent MME approaches, including arithmetic mean ensemble (ENS), multiple linear regression (MLR), support vector machine (SVM), random forest (RF), and K-nearest neighbors (KNN), against in-situ measurements during 1970–2014. Overall, EC_Earth3_Veg outperforms other GCMs and ENS, with most monthly taylor skill scores (TSSs) between 0.8 and 0.9. Its performance is further improved through QM bias-correction, which however still presents slight gaps compared to other MME methods except SVM. CNN performs best with the highest comprehensive rating metric (MR) value of 0.96, corresponding to the highest mean TSS of 0.98, followed by KNN, RF, and MLR. It is also optimal for reproducing geographic distributions and seasonal cycle of precipitation and reducing biases at local scale, attributed to its unique advantage in resolving spatial dependences. Despite being slightly inferior to the best GCM, it outperforms other MME methods in capturing observed interannual variations and probability density functions, with weak sensitivity to varying ensemble sizes. Overall, CNN model is highly recommended in multi-GCM ensemble predictions for its superior and stable performance.