We investigate how state-of-the-art Large Language Models (LLMs) can unlock new knowledge from the Cyfrowa Biblioteka Druków Ulotnych (CBDU)—a digital collection of early-modern Polish ephemeral prints. Our end-to-end pipeline compares three transcription approaches: pure OCR, LLM-based post-correction, and multimodal models. The resulting transcriptions then serve as input for our two main contributions: the automatic extraction of bibliographic metadata and the generation of expert-style historical commentaries. Experiments show a leading multimodal model excels, reducing transcription CER from 33% to 9%, while achieving high F1-scores for publication place (0.85) and date (0.71), and a 2.31/3 mean score for commentaries. We conclude that large multimodal models can serve as effective “digital archivists”, enriching historical collections with structured metadata and contextual analysis.