How to Run On-Device LLM Summarization in Android Background Using WorkManager

A developer tutorial outlines a pattern for integrating Android's WorkManager with a quantized on-device large language model — specifically llama.cpp via JNI — to perform background document summarization reliably. The architecture uses chained Workers in a parallel fan-out and serial reduce pattern, processing documents in roughly 500-token chunks to avoid memory overflows and the 10-minute execution window limit. Model selection is tied strictly to device memory ceilings, with mid-range 6 GB devices recommended to use only sub-1.5B INT4 quantized models for background tasks without foreground promotion. Key WorkManager constraints include setting battery-not-low requirements and using setExpedited to prevent Doze-mode from deferring inference jobs by hours. The guide targets Android API 26+ with WorkManager 2.9+ and applies broadly to llama.cpp, MLC LLM, or MediaPipe LLM Inference API integrations.