MobileSpec trains 8B LLM specialists on iPhone overnight

Researchers present MobileSpec, a system that trains specialized large language models on personal devices using forward passes alone and learning from automatically checked task outcomes. The core algorithm, FuRA-S, updates block-spectral coordinates of pretrained weights to produce efficient perturbations similar to reinforcement learning with verifiable rewards. The approach enables training a personalized 8-billion-parameter model on an iPhone overnight while substantially improving phone-use agent success rates. Open article →