Coronal Mass Ejections (CMEs) are the primary source of strong space weather disturbances at Earth. Their geoeffectiveness is largely determined by their dynamic pressure and internal magnetic fields, for which reliable predictions at Earth are not possible with traditional cone CME models. We study two Earth-directed CMEs using the EUropean Heliospheric FORecasting Information Asset (EUHFORIA) model, testing the predictive capabilities of a linear force-free spheromak CME model initialised using parameters derived from remote-sensing observations. Using observation-based CME input parameters, we perform MHD simulations of the events with the cone and spheromak CME models. Results show that spheromak CMEs propagate faster than cone CMEs when initialised with the same kinematic parameters. We interpret these differences as due to different Lorentz forces acting within cone and spheromak CMEs, leading to different CME expansions. Such discrepancies can be mitigated by initialising spheromak CMEs with a reduced speed corresponding to the radial speed only. Results at Earth evidence that the spheromak model improves the predictions of B(Bz) up to 12-60(22-40) percentage points compared to a cone model. Considering virtual spacecraft located around Earth, B(Bz) predictions reach 45-70%(58-78%) of the observed peak values. The spheromak model predicts inaccurate magnetic field parameters at Earth for CMEs propagating away from the Sun-Earth line, while it successfully predicts the CME properties and arrival time in the case of strictly Earth-directed events. Modelling CMEs propagating away from the Sun-Earth line requires extra care due to limitations related to the assumed spherical shape. The spatial variability of modelling results and the typical uncertainties in the reconstructed CME direction advocate the need to consider predictions at Earth and at virtual spacecraft around it.