Simulations based on disease models can be used as a training tool for patient education or caregiver support to improve effectiveness of practice. Disease models provide predictions of patient outcomes, costs, and quality of life information over time. They are typically constructed by various teams around the world based on local data, and currently those do not validate well against multiple external populations. Therefore, universal understanding of disease progression is still unsolved. The Reference Model for Disease Progression addresses this problem by implementing a league of disease models that compete for fitness towards publicly available clinical data. Currently, diabetic populations are of interest and data for the model is drawn from published clinical trials, while model building blocks are based on published risk equations and modeling hypotheses. The Reference Model creates model combinations from those building blocks, then simulates and validates them against multiple populations. High Performance Computing (HPC) techniques are used to cope with the combinatorial number of models that, until recently, took roughly a year of computation time on a single processor. Recently new building blocks have been added to the model, and the number of combinations became too large to compute in reasonable time without access to a large cluster. To cope with this, the structure of the model was changed from a competitive discrete ensemble model, where building blocks are selected from a discrete pool of options to construct the best model, to a cooperative continuous ensemble model, where all building blocks are merged using linear combination. Moving the model toward continuous combination space allowed creation of an assumption engine that employs optimization algorithms to better deduce the most fitting model. This significantly reduces simulation time and produces a better fitting model. This paper will focus on recent modifications, and results obtained from the latest simulations.