We propose a unified framework for model merging that leverages multiple symmetry classes to enable low- and zero-loss interpolation between independently trained Transformer models, including Vision Transformers and GPT-2.