The size of grains delivered to rivers by hillslope processes is thought to be a key factor controlling sediment transport, long-term erosion and the information recorded in sedimentary archives. Recently, models have been developed to estimate the grain size distribution produced in soil, but these models may not apply to active orogens where high erosion rates on hillslopes are driven by landsliding. To date, relatively few studies have focused on landslide grain size distributions. Here, we present grain size distributions (GSDs) obtained by grid-by-number sampling on 17 recent landslide deposits in Taiwan, and we compare these GSDs to the geometrical and physical properties of the landslides, such as their width, area, rock type, drop height and estimated scar depth. All slides occurred in slightly metamorphosed sedimentary units, except two, which occurred in younger unmetamorphosed shales, with a rock strength that is expected to be 3–10 times weaker than their metamorphosed counterparts. For 11 landslides, we did not observe substantial spatial variations in the GSD over the deposit. However, four landslides displayed a strong grain size segregation on their deposit, with the overall GSD of the downslope toe sectors being 3–10 times coarser than apex sectors. In three cases, we could also measure the GSD inside incised sectors of the landslides deposits, which presented percentiles that were 3–10 times finer than the surface of the deposit. Both observations could be due to either kinetic sieving or deposit reworking after the landslide failure, but we cannot explain why only some deposits had strong segregation. Averaging this spatial variability, we found the median grain size of the deposits to be strongly negatively correlated with drop height, scar width and depth. However, previous work suggests that regolith particles and bedrock blocks should coarsen with increasing depth, which is the inverse of our observations. Accounting for a model of regolith coarsening with depth, we found that the ratio of the estimated original bedrock block size to the deposit median grain size (D50) of the deposit was proportional to the potential energy of the landslide normalized to its bedrock strength. Thus, the studied landslides agree well with a published, simple fragmentation model, even if that model was calibrated on rock avalanches with larger volume and stronger bedrock than those featured in our dataset. Therefore, this scaling may serve for future modeling of grain size transfer from hillslopes to rivers, with the aim to better understanding landslide sediment evacuation and coupling to river erosional dynamics.