Ensemble forecast post-processing techniques are often applied to improve the reliability of the weather forecast. These methods issue a future predictive distribution which is better calibrated and as sharp as possible. While the field of forecast post-processing has been rapidly advancing, especially with the advent of machine learning, many challenges remain: post-processing models often assume the target distribution or require specialised distributions for each target variable, they require separate models for each lead time or location, and frequently forgo the explicit modelling of the temporal correlation structure in the target variable. In this work, we aim to address the outlined shortcomings by proposing a novel post-processing method which combines normalising spline flows and flow matching. This combination allows us to efficiently estimate marginal distributions and generate temporally correlated post-processed forecasts, useful for downstream applications. We evaluate our model on the openly available EUPPBench dataset, post-processing three target weather variables: temperature, precipitation, and wind gusts. Our method achieves competitive results for both marginal and joint distribution post-processing for all considered weather variables, rivalling current state-of-the-art techniques. Furthermore, we show that our joint lead-time post-processing approach extends its performance margin in a predictor-limited environment compared to per-lead-time methods.