Local seismic networks are usually designed so that earthquakes are located inside them (primary azimuthal gap <<180°) and close to the seismic stations (0-100 km). With these local or near-regional networks (0°-5°), many seismological observatories still routinely locate earthquakes using 1-D velocity models. Moving towards 3-D location algorithms requires robust 3-D velocity models. This work takes advantage of seismic monitoring spanning more than 30 yr in the Pyrenean region. We investigate the influence of a well-designed 3-D model with station corrections including basins structure and the geometry of the Mohorovicic discontinuity on earthquake locations. In the most favourable cases (GAP < 180° and distance to the first station lower than 15 km), results using 1-D velocity models are very similar to 3-D results. The horizontal accuracy in the 1-D case can be higher than in the 3-D case if lateral variations in the structure are not properly resolved. Depth is systematically better resolved in the 3-D model even on the boundaries of the seismic network (GAP > 180° and distance to the first station higher than 15 km). Errors on velocity models and accuracy of absolute earthquake locations are assessed based on a reference data set made of active seismic, quarry blasts and passive temporary experiments. Solutions and uncertainties are estimated using the probabilistic approach of the NonLinLoc (NLLoc) software based on Equal Differential Time. Some updates have been added to NLLoc to better focus on the final solution (outlier exclusion, multiscale grid search, S-phases weighting). Errors in the probabilistic approach are defined to take into account errors on velocity models and on arrival times. The seismicity in the final 3-D catalogue is located with a horizontal uncertainty of about 2.0 ± 1.9 km and a vertical uncertainty of about 3.0 ± 2.0 km.