Mobile LLM applications face a cold-start problem: the first token in a conversational turn often arrives 800–1200ms after the user finishes speaking, even with optimized models. Weight loading is solved via memory-mapping; tokenization takes