We discuss how to avoid aggregation bias in large-scale global Computable General Equilibrium (CGE) models by reducing the need of pre-model aggregation, based on the combination of algorithmic improvements and a filtering approach which removes small transactions. Using large-scale sensitivity analysis, we show the impact of pre-aggregation and filtering on model size, model solution time and simulated welfare impacts, using a multi-lateral partial trade liberalization simulated with the standard GTAP model as the test case. We conclude that pre-model aggregation should be avoided as far as possible, and that our filtering approach and algorithmic improvements allow global CGE analysis even with highly disaggregated data sets at moderate solution times.