<p>The International Association of Geodesy defined the International Height Reference System (IHRS) in 2015. IHRS will be realised through the International Height Reference Frame (IHRF) primarily by geopotential numbers. On the global scale, a sparse core reference network is utilised to realise the IHRS, but denser realisations (or densifications) are needed to ensure local access and height datum unification. ITRF ellipsoidal heights and a gravimetric (quasi-) geoid model form the basis for a model-based IHRS realisation, which can be enhanced by levelling to improve local accessibility to the global frame and relative uncertainty. This study investigates incorporating levelling observations in the IHRS realisation process and explores potential improvements. A case study over Sweden using an initial model-based IHRS realisation and precise levelling observations from the Baltic Levelling Ring (BLR) project was performed. The levelling network was adjusted relative to the model-based IHRF geopotential numbers using three approaches regarding the a priori variance–covariance matrix for weighting of the observations: constrained adjustment, weighted adjustment with realistic a priori uncertainties, and weighted adjustment with Variance Component Estimation (VCE) to iteratively tune the variance–covariance matrix. It is shown that the model-based IHRS realisation improves significantly. Both the estimated standard uncertainty of the adjusted IHRF geopotential numbers and the relative standard uncertainty between nodal benchmarks are reduced by approximately 40 per cent in the Swedish case. To fully exploit precise levelling observations in the IHRS realisation process, a weighted adjustment, preferably using VCE to optimally combine the initial model-based geopotential numbers and the levelling observations is recommended.</p>