Generative AI, and particularly language models (LLM), have sparked a new revolution in AI with applications across all domains. However, LLMs are highly resource-intensive and, hence, difficult to implement on autonomous embedded systems. LLMs can be optimized by modifying their architecture to replace heavy Transformer layers with lighter alternatives. Given the difficulty of training LLM "from scratch," this thesis aims to develop post-training neural architecture optimization methods applicable to small LLM (SLM). Additionally, the thesis seeks to propose performance metrics of different layers of an SLM and their alternatives, to guide the replacement, and thus propose a comprehensive methodology for optimizing SLMs while considering hardware constraints. The work will be valorized through publications in major AI conferences and journals, and the developed codes and methods could be integrated into the tools developed at CEA.