LLMP6: A six-layer composable architecture for large language model applications

This paper proposes LLMP6, a six-layer composable architecture designed for large language model applications. The architecture decomposes the LLM application workflow into six standardized layers—Assign, Core, Tool, Filter, Check, and Show—each supporting independent model selection. The Core layer utilizes a foundational large model, while the remaining five layers are implemented with lightweight specialized small models. Key innovations include heterogeneous model collaboration, parallel/serial execution modes, automatic optimization mechanisms, and core iteration processes. Experimental results demonstrate that under the LLMP6-Full configuration, the response times of DeepSeek and Kimi were reduced by 44.7% and 17.2%, respectively; throughput increased by 86.2% and 21.4%; P99 latency decreased by 77.4% and 51.0%; latency jitter decreased by 74.7% and 38.0%; the coefficient of variation decreased by 64.3% and 38.3%; DeepSeek's F1 score improved from 77.6% to 79.8%, while Kimi's F1 score rose from 68.5% to 69.8%. Changes in accuracy and quality scores remained below 1.5%, with output diversity and readability maintained consistently, achieving an optimal balance between efficiency and quality.