Key innovations in Meta’s MobileLLM include prioritizing model depth over width, implementing embedding sharing and grouped-query attention and utilizing a novel immediate block-wise weight-sharing technique.