Federated Large Language Models (LLMs) are emerging as an essential technology in education, enabling personalized learning experiences while preserving data privacy. However, deploying LLMs in school settings introduces several challenges, including limited computational resources and the need to coordinate learning across diverse institutions. This proposal explores the integration of federated learning frameworks such as FATE and Flower , investigating whether homogeneous or heterogeneous environments, knowledge distillation, or parameter-efficient tuning techniques (like LoRA) yield better performance. We aim to establish a scalable system where smaller edge LLMs, like SMOLLM , interact with larger backbone models such as LLaMA 70B. Our goal is to identify optimal strategies for deploying federated LLMs in educational environments, balancing privacy, performance, and scalability. Motivation and Goal Federated Learning (FL) in education can transform how schools utilize AI, offering personalized learning without compromising data privacy. In a federated setup, multiple schools can collaborate on training LLMs while keeping student data local. This approach is especially relevant in a time where privacy regulations are stringent, and the diversity of educational environments requires adaptable AI solutions. The use of FATE-LLM and Flower frameworks enables this by supporting both homogeneous (identical models across schools) and heterogeneous (diverse models) setups. Our exploration focuses on several critical aspects: • Edge vs. Centralized Models: Integrating lightweight edge models like SMOLLM with large server-side models such as LLaMA 70B for better adaptability and resource management. • Fine-Tuning Strategies: Evaluating different methods like LoRA, full fine-tuning, and knowledge distillation to optimize performance. • Heterogeneous vs. Homogeneous Environments: Understanding the impact of varied model sizes and architectures on the overall system efficiency. This project aims to assess which configuration of federated LLMs best suits educational contexts, with experiments designed to measure both performance and scalability.