There is great interest in developing chemical kinetic models with predictive power. A predictive kinetic model maintains high fidelity to the true chemistry usually via a full-detail molecule representation which distinguishes all the molecules and isomers. Unfortunately such full-detail representation quickly becomes intractable for large molecules because the number of isomers grows very rapidly with size. Several less detailed representations have been developed, including the popular Structure-Oriented Lumping (SOL) method, which allow one to model larger molecules. In this paper we first examine scalability for two existing representations (full-detail and SOL) and propose a new modeling methodology tailored for heavy systems: fragment-based modeling, which promises the best scalability. We show via a case study of heavy molecule pyrolysis that the new approach creates a much smaller reaction network but with similar prediction accuracy on feedstock conversion and products' molecular weight distribution compared to the corresponding model using full-detail representation. An open source modeling package AutoFragmentModeling using this methodology has been developed. Some suggestions are presented for future work based on this new modeling approach.