<jats:p>Abstract. 10Be ice core measurements are an important tool for paleoclimate research, e.g., allowing for the reconstruction of past solar activity or changes in the geomagnetic dipole field. However, especially on multi-millennial timescales, the share of production and climate-induced variations of respective 10Be ice core records is still up for debate. Here we present the first quantitative climatological model of the 10Be ice concentration up to the glacial–interglacial timescale. The model approach is composed of (i) a coarse resolution global atmospheric transport model and (ii) a local 10Be air–firn transfer model. Extensive global-scale observational data of short-lived radionuclides as well as new polar 10Be snow-pit measurements are used for model calibration and validation. Being specifically configured for 10Be in polar ice, this tool thus allows for a straightforward investigation of production- and non-production-related modulation of this nuclide. We find that the polar 10Be ice concentration does not immediately record the globally mixed cosmogenic production signal. Using geomagnetic modulation and revised Greenland snow accumulation rate changes as model input, we simulate the observed Greenland Summit (GRIP and GISP2) 10Be ice core records over the last 75 kyr (on the GICC05modelext timescale). We show that our basic model is capable of reproducing the largest portion of the observed 10Be changes. However, model–measurement differences exhibit multi-millennial trends (differences up to 87% in case of normalized to the Holocene records) which call for closer investigation. Focusing on the (12–37) b2k (before the year AD 2000) period, mean model–measurement differences of 30% cannot be attributed to production changes. However, unconsidered climate-induced changes could likely explain the model–measurement mismatch. In fact, the 10Be ice concentration is very sensitive to snow accumulation changes. Here the reconstructed Greenland Summit (GRIP) snow accumulation rate record would require revision of +28% to solely account for the (12–37) b2k model–measurement differences. </jats:p>