Meta AI has unveiled a new paper, MobileLLM, focusing on optimizing sub-billion parameter language models for on-device use cases. The key innovations in MobileLLM include prioritizing model depth over width, shared matrices for token-to-embedding and embedding-to-token, and shared weights between multiple transformer blocks. This development aims to make smaller language models suitable for smartphones and other edge devices, potentially saving high cloud costs and improving response times. Meta has released the source code of MobileLLM, including results for models ranging from 125M to 1.5B parameters. The move from large-scale models to more compact ones is seen as a new trend in AI, making the technology more accessible to billions of people. MobileLLM was also presented at ICML 2024.