Research on An Independent Game Intelligent NPC Adaptation Solution Based on Lightweight LLMs—Taking "Ghost Valley: Eight Wastelands" as An Example

Independent game development has long been constrained by the rigidity of NPC interaction models, while haphazardly introducing heavy, large language models (LLMs) faces the practical barrier of excessively high computational costs. To resolve this engineering dilemma, this paper constructs and implements a lightweight LLM-based intelligent NPC adaptation framework. Using the cultivation-themed game "Ghost Valley: Eight Directions" as a testbed, the study integrates the DeepSeek-r1-14B model at the core layer and locally deploys it with INT4 quantization technology, fully establishing the decoupled three-tier chain of "model-API-engine." Test data demonstrates that this solution runs smoothly on ordinary consumer-grade PCs. It not only maintains the logical coherence of NPC dialogue and character consistency but also reduces memory usage and inference latency to levels acceptable for independent teams. Starting from engineering practice, this research provides a cost-effective Artificial Intelligence(AI) implementation reference for development teams with limited computational resources.