The Arctic is one of Earth’s regions highly susceptible to climate change. However, in situ long-term observations used for climate research are relatively sparse in the Arctic Ocean, and current climate models exhibit notable biases in Arctic Ocean simulations. Here, we present an Arctic Ocean dynamical downscaling dataset, obtained from the global ocean-sea ice model FESOM2 with a regionally refined horizonal resolution of 4.5 km in the Arctic region, which is driven by bias-corrected surface forcings derived from a climate model. The dataset includes 115 years (1900–2014) of historical simulations and two 86-year future projection simulations (2015–2100) for the SSP2-4.5 and SSP5-8.5 scenarios. The historical simulations demonstrate substantially reduced biases in temperature, salinity and sea-ice thickness compared to CMIP6 climate models. Common biases in the representation of the Atlantic Water layer found in climate model simulations are also markedly reduced in the dataset. Serving as a crucial long-term data source for climate change assessments and scientific research for the Arctic Ocean, this dataset provides valuable information for the scientific community.