The performance of text recognition models, whether OCR or HTR, is improving each year, with a recent acceleration thanks to hybrid models that combine vision and language models. These systems are typically based on well-controlled data: the languages covered have few variations, and efficient post-correction mechanisms can be applied thanks to language models. However, when it comes to medieval manuscripts, the situation becomes significantly more complex. These documents exhibit much more heterogeneous characteristics, which challenge current models that rely heavily on their ability to understand standardized and regular language.