We present redshift distribution estimates of galaxies selected from the fourth data release of the Kilo-Degree Survey over an area of ∼1000 deg² (KiDS-1000). These redshift distributions represent one of the crucial ingredients for weak gravitational lensing measurements with the KiDS-1000 data. The primary estimate is based on deep spectroscopic reference catalogues that are re-weighted with the help of a self-organising map (SOM) to closely resemble the KiDS-1000 sources, split into five tomographic redshift bins in the photometric redshift range 0.1 < z_B ≤ 1.2. Sources are selected such that they only occupy that volume of nine-dimensional magnitude-space that is also covered by the reference samples (‘gold’ selection). Residual biases in the mean redshifts determined from this calibration are estimated from mock catalogues to be ≲0.01 for all five bins with uncertainties of ∼0.01. This primary SOM estimate of the KiDS-1000 redshift distributions is complemented with an independent clustering redshift approach. After validation of the clustering-z on the same mock catalogues and a careful assessment of systematic errors, we find no significant bias of the SOM redshift distributions with respect to the clustering-z measurements. The SOM redshift distributions re-calibrated by the clustering-z represent an alternative calibration of the redshift distributions with only slightly larger uncertainties in the mean redshifts of ∼0.01 − 0.02 to be used in KiDS-1000 cosmological weak lensing analyses. As this includes the SOM uncertainty, clustering-z are shown to be fully competitive on KiDS-1000 data.