Enabling Local Multimodal AI for Metadata Generation in Archival Collections

Abstract. Large-scale digitisation has transformed access to archives, but generating reliable metadata remains a bottleneck. This short study explores whether locally deployed open-source multimodal models can assist archivists by generating descriptive metadata based on the “Four Ws” (Who, What, Where, When). We implement a proof-of-concept using Qwen, supported by GPT-assisted curation of Winston Churchill images, and evaluate its potential for integration into digital catalogues.