This thesis investigates whether general-purpose Multimodal Large Language Models can serve as reliable tools for historical information extraction within Digital Humanities background. Using a sample of 100 Swedish historical patent cards (1945–1975), the study evaluates both proprietary and open-source MLLMs through a pipeline adapted from prior work by Xie, Mela and Tell (2025). Beyond performance metrics, the analysis is framed by four core values in DH: accessibility, transparency and reproducibility, iterative design, and ethical responsibility.

Proprietary models – especially Gemini 2.5 Pro – demonstrated strong accuracy across multiple types of text transcription and showed reliable potential in handling reasoning tasks. However, cloud-based platforms continue to raise concerns regarding the opacity of training data, the privacy of user data and the bias hidden in the model structure. In contrast, locally deployed open-source models, such as Qwen2.5-VL, offer data security and greater control over input processing. Yet, under standard consumer-grade hardware configurations, they proved limited in performance. By embedding evaluation within a real research workflow and comparing deployment scenarios, this study proposes a practical, value-driven framework for selecting and assessing AI tools in DH contexts. The findings highlight the importance of critical tool engagement and demonstrate how MLLMs, when thoughtfully implemented, can support scalable, transparent, and ethically sound digital scholarship.