Practice of Mobile Application LLM System Driven by End-Edge-Cloud Collaboration

Abstract

Large language models (LLMs) have rapidly become a general-purpose capability layer for mobile applications, yet "cloud-only LLM" deployment faces persistent bottlenecks in privacy-sensitive data access, inference cost, and real-time reliability. This paper presents a practical system design for a mobile application LLM stack driven by end-edge-cloud collaboration and coordinated "large-small model" execution. We summarize why a single large model cannot adequately address (i) user-level data richness and privacy constraints on-device, (ii) the high marginal cost of cloud inference at scale, and (iii) responsiveness and stability requirements under variable networks. We propose an architecture that assigns personalized, latency-critical, and privacy-preserving functions to on-device small models and local runtimes; delegates cacheable, low-latency coordination and retrieval services to the edge; and reserves cloud LLMs for complex reasoning and generation. We further describe orchestration mechanisms, routing policies, and optimization techniques, including context condensation, selective retrieval, speculative execution, and feedback-driven adaptation. Results are reported as system-level outcomes in terms of latency, cloud token reduction, and robustness under network degradation, concluding that end-edge-cloud collaboration can improve user experience while materially reducing cloud-side cost and expanding privacy-respecting capability coverage.