Arctic sea ice has substantially declined over the past four decades. Given the projected decrease under global warming, reliable prediction of summer sea ice is becoming increasingly important. Yet most climate models still show limited low skill in predicting summer sea ice from spring, a limitation commonly referred to as the spring predictability barrier. Here we develop a seasonal forecast system based on an eddy-permitting coupled model and assess how different ocean and sea ice initialization strategies affect prediction skill. Results show that the summer sea ice in the Barents-Kara Sea is predicted most skillfully when sea ice and snow conditions in the model are initialized. Sea ice initialization improves the representation of the initial sea ice thickness, the subsequent ice-albedo feedback, and the meridional sea ice advection from the higher latitudes. These findings suggest significance of sea ice thickness initialization for skillful summer prediction despite the spring predictability barrier.