This study compares the skills of two numerical models at the same horizontal resolution but based on different principles in representing meso- and sub-mesoscale ocean features. The first model, titled LD20-NEMO, was based on solving primitive equations of ocean dynamics. The second model, titled LD20-SDD, used a newer stochastic–deterministic downscaling (SDD) method. Both models had 1/20° resolution, the same meteo forcing, and used outputs from a data assimilating global model at 1/12° resolution available from Copernicus Marine Service (CMEMS). The LD20 models did not assimilate observational data but were physically aware of observations via the parent model. The LD20-NEMO only used a 2D set of data from CMEMS as the lateral boundary conditions. The LD20-SDD consumed the full 3D set of data from CMEMS and exploited the stochastic properties of these data to generate the downscaled field variables at higher resolution than the parent model. The skills of the three models were assessed against remotely sensed and in situ observations for the four-year period 2015–2018. The models showed similar skills in reproducing temperature and salinity, however the SDD version performed slightly better than the NEMO, and was more computationally efficient by a large margin.