Model merging is a powerful technique to combine specialized knowledge of multiple machine learning models into a single unified model.However, current methods require manually partitioning the model parameters into a fixed number of groups to be merged, which constraints the exploration of potential combinations and limits performance.To address these limitations, we propose an evolutionary algorithm with three key features:(1) dynamically adjustment of merging boundaries to progressively explore a broader range of parameter combinations;(2) a diversity preservation mechanism inspired by nature, which maintains a population of diverse, high-performing models that are particularly effective for merging;and (3) a heuristic-based \textit{mate selection} strategy to identify the most promising pairs of models for merging. Our experimental results show, for the first time, that model merging can be used to evolve models from \textit{scratch}.Specifically, we evolve MNIST classifiers from scratch using our method, and achieve comparable performance to CMA-ES, while being computationally cheaper.Additionally, we use our method to merge specialised language models and obtain state-of-the-art performance.Our code is available at https://github.com/AnonScientist/natural_niches.