The Final Hybrid: Merging Two Monolingual Language Models Through Genetic Tensor Expansion Without Retraining

This work documents the successful hybridization of two monolingual language models without retraining. Each model was trained on a different single language. The method uses zero-padding expansion for 2D tensors and repetition for 1D tensors. RAM is limited to 4GB. The tokenizer formats differ (tokenizer.json vs vocab.json) and require manual alignment before hybridization. Results confirm that both original languages are preserved, no single model dominates, and emergent cross-lingual capabilities are observed. The hybrid model can evolve further through additional hybridizations. Code is available on GitHub: https://github.com/mouad-tarif/Monolingual-Hybridization