Background Excess mortality from all-cause has been estimated at national level for different countries, to provide a picture of the total burden of the COVID-19 pandemic. Nevertheless, there have been no attempts at modelling it at high spatial resolution, needed to understand geographical differences in the mortality patterns, to evaluate temporal lags and to plan for future waves of the pandemic. Methods This is the first subnational study on excess mortality during the COVID-19 pandemic in Italy, the third most-hit country. We considered municipality level and estimated all-cause mortality weekly trends based on the first four months of 2016 – 2019. We specified a Bayesian hierarchical model allowing for spatial heterogeneity as well as for non-linear smooth spatio-temporal terms. We predicted the weekly mortality rates at municipality level for 2020 based on the modelled spatio-temporal trends (i.e.~in the absence of the pandemic) and estimated the excess mortality and the uncertainty surrounding it. Results There was strong evidence of excess mortality for Northern Italy; Lombardia showed higher mortality rates than expected from the end of February, with 23,946 (23,013 to 24,786) total excess deaths. North-West and North-East regions showed higher mortality from the beginning of March, with 6,942 (6,142 to 7,667) and 8,033 (7,061 to 9,044) total excess deaths respectively. After discounting for the number of COVID-19-confirmed deaths, Lombardia still registered 10,197 (9,264 to 11,037) excess deaths, while regions in the North-West and North-East had 2,572 (1,772 to 3,297) and 2,047 (1,075 to 3,058) extra deaths, respectively. We observed marked geographical differences at municipality level. The city of Bergamo (Lombardia) showed the largest percent excess 88.9% (81.9% to 95.2%) at the peak of the pandemic. An excess of 84.2% (73.8% to 93.4%) was also estimated at the same time for the city of Pesaro (Central Italy), in stark contrast with the rest of the region, which does not show evidence of excess deaths.