Abstract. Fine particulate matter (PM2.5) has altered radiation balance on earth and raised environmental and health risks for decades, but only been monitored widely since 2013 in China. Historical long-term PM2.5 records with high temporal resolution are essential but lacking for both research and environmental management. Here, we reconstruct a site-based PM2.5 dataset at 6-hour intervals from 1960 to 2020 that combines long-term visibility, conventional meteorological observations, emissions, and elevation. The PM2.5 concentration at each site is estimated based on an advanced machine learning model, LightGBM, that takes advantage of spatial features from 20 surrounding meteorological stations. Our model's performance is comparable or even better than those of previous studies in by-year cross validation (CV) (R2=0.7) and spatial CV (R2=0.76), and more advantageous in long-term records and high temporal resolution. This model also reconstructs a 0.25°×0.25°, 6-hourly, gridded PM2.5 dataset by incorporating spatial features. The results show PM2.5 pollution worsens gradually or maintains before 2010 from an interdecadal scale but mitigates in the following decade. Although the turning points vary in different regions, PM2.5 mass concentrations in key regions decreased significantly after 2013 due to clean air actions. In particular, the annual average value of PM2.5 in 2020 is nearly at the lowest value in history since 1960. These two PM2.5 datasets (publicly available at https://doi.org/10.5281/zenodo.6372847) provide spatiotemporal variations at high resolution, which lay the foundation of research studies associated with air pollution, climate change, and atmospheric chemical reanalysis.