Effekte der Quantisierung auf Retrieval-Augmented-Generation Language Modelle für die Parkinson-Versorgung

Large Language Models hold significant potential for assistive applications in Parkinson’s care, yet their substantial memory requirements currently limit deployment in real-world clinical environments. This paper investigates the impact of different quantization methods on a Retrieval-Augmented Generation system based on a Qwen model. Our results show that model size can be reduced to as little as 36% of the original footprint, with only mild to moderate degradation in response quality. These findings demonstrate that locally executable, Parkinson's-specific assistant systems are technically feasible and enable privacy-preserving assistance systems.