Dual-Model Simple Cache (DMSC) and Triple-Model Adaptive Routing (TMAR): Multi-Model Optimization

Abstract: This paper introduces two innovative approaches for optimizing large language model query sys-
tems: Dual-Model Simple Cache (DMSC) and Triple-Model Adaptive Routing (TMAR). DMSC achieves
a 93.4% reduction in response time by implementing an intelligent caching mechanism using Redis, com-
bining a lightweight cache model (Gemma 2 9B) with a powerful main model (Deepseek R1) to optimize
repeated queries. TMAR delivers a 35.8% reduction in average response time through complexity-based
routing, utilizing a three-tiered architecture that intelligently directs questions based on their complexity
level. Both systems demonstrate significant performance improvements while maintaining response quality,
offering practical solutions for efficient large language model deployment.