Deep fractured-vuggy carbonate reservoirs, shaped by paleogeomorphology, structures, and karstification, exhibit various types, significant scale differences, high discreteness, and extreme heterogeneity, making it difficult to immediately apply conventional modelling methods to these reservoirs. In this study, we investigate the Ordovician carbonate reservoirs in the Tahe oilfield. To deal with the key issues in modelling these reservoirs, including a lack of guidance on the reservoirs’ developmental models, constraints from actual statistical laws, and the modelling and optimization algorithms of fracture-cave architectures, we develop a series of fine modelling techniques for the fracture-vug architectures, which center on genetic classification, multiple constraints, and multi-point statistics. Regarding reservoir architecture characterization, physical property modelling, and multi-type model integration, we introduce five major distinctive techniques with great efforts, namely multiple-point statistical modelling for a subsurface paleo-river system, multi-constraint modelling for fault-controlled karst reservoirs under zoning, collaborative simulation for genesis-controlled epikarst reservoirs, physical property simulation using karst facies control combined with equivalent calculation, and genetic sequence-based multi-type model integration and relevant dynamic optimization. The geological modelling of fractured-vuggy reservoirs has undergone four shifts: a shift in the modelling objects from external reservoir morphology to the internal geneses architecture of fractures, pores and caves, a shift in inter-well reservoir simulation from a single constraint to multiple control, a shift in modelling methods from the predominance of seismically sculptured geological modelling to multidisciplinary collaborative characterization, and a shift in modelling tools from merely commercial software to a combination of commercial software and independently developed modules. The results indicate that the techniques developed in this study can significantly enhance the coincidence rate of the geological model developed with drilled wells, serving to support geological modelling of 15 units in the Tahe oilfield, covering geological reserves of approximately 1.5×108 t. Additionally, the newly developed model has been employed to refine reserve composition, perform numerical simulation of reservoirs, and adjust development schemes with remarkable effects, laying a geological foundation for developing measures to tap into the remaining oil potential and enhance oil recovery of fractured-vuggy carbonate oil reservoirs.