MobileLLM provides training code and models for compact language models under one billion parameters, using design choices like SwiGLU activation, deep thin architectures, embedding sharing, and grouped-query attention to maximize quality at small scale. The project has released multiple iterations including MobileLLM-R1 and MobileLLM-R1.5, with models publicly available on HuggingFace and validation across math, coding, and commonsense reasoning benchmarks.