Large Language Models typically perform best when there is a lot of data to train on. An interesting question is how such models perform in less dominant languages (e.g., Swedish) and over domains of historical interest where less modern words and phrasings are used. Specifically, how can existing language models for current usage be adapted to get the best question-answering results when applied over such historical domains? In this Master's thesis, we will use domain adaption on three different pre-trained Swedish NLP language models: BERT base Swedish Cased SQuAD Experimental, GPT-SW3, and IBM Watson. We will fine-tune these models on the specific domain of Vasaloppet's 100-year-old articles to evaluate if the models will have difficulties predicting results from a Swedish source that is a century old. We will evaluate how well these models perform on question answering and report results and lessons learned.