This study investigates the performance of the elliptic blending differential flux model (EBDFM) and the standard differential flux model (DFM) in reproducing turbulent natural convection in a square cavity. In both cases, the Reynolds stresses are computed with the Elliptic Blending Reynolds Stress models (EBRSM). The model coefficient associated with the production of dissipation are modified to promote turbulence in the boundary layer, without which the velocity is overestimated and boundary layer is too thin. The computed mean quantities are in good agreement with data from the experiment. The vertical velocity component, the mean temperature and the local Nusselt number along the walls are reproduced with good accuracy. EBDFM show improved prediction of the mean horizontal velocity component, although the numerical predictions are not very accurate. In, general the EBDFM produces a better prediction of the turbulent quantities. While both models overpredict the peaks of the turbulent fields, the EBDFM is closer to the experiment. The DFM also overpredicts the turbulent vertical components in the outer boundary layer. Numerical predictions of the horizontal component of turbulent heat flux show weak accuracy with both models overestimating the peak by about three times. However, despite some discrepancies, the EBDFM shows improved performance over the DFM in the reproduction of the turbulent quantities and proves the advantage of the elliptic blending approach in modelling near wall turbulence.