Different sets of forcing data and types of hydrological models (particularly data-driven and physics-based models) with their respective inherent errors can lead to varying systematic deviations in streamflow simulations. These systematic errors might be mitigated through the use of ensemble simulations. To this end, we first trained and calibrated six models on 671 basins across the contiguous United States (CONUS) using three meteorological forcing datasets and two types of advanced models: a differentiable HBV (Hydrologiska Byråns Vattenbalansavdelning) model and a deep learning LSTM (Long Short-Term Memory) model. Consequently, the independent deviations caused by different forcing datasets and model types were generated and comprehensively assessed. We then investigated the simulation differences and combined the results using optimized weights to determine if independent errors could be reduced, which were tested through temporal, PUB (prediction in ungauged basins), and PUR (prediction in ungauged regions) tests. The results consistently showed significant improvements in streamflow prediction across all three tests and set new state-of-the-art records on the 671 basins. This study demonstrates the effectiveness of the ensemble approach via both differentiable HBV and LSTM in reducing independent errors caused by varying forcing datasets and model types.