Recently, many educators have been using generative AI to create materials for their courses. However, they often report frustration with the inability of Large Language Models (LLMs) to reliably produce language appropriate to their students’ proficiency levels. This raises the question: Can LLMs be adapted to generate level-appropriate learning materials consistently? This research project aims to develop an LLM capable of producing output at each CEFR level by taking into account vocabulary, grammatical features, and lexical complexity. Three open-source LLMs were selected and fine-tuned using datasets of level-differentiated CEFR texts. This presentation explains the fine-tuning process and compares the effects of different datasets on the three models. It will also introduce tools and workflows that allow participants to fine-tune models to suit their own need. Model outputs at each level will be evaluated against CEFR benchmarks, and the output of each model compared and shared with the participants for discussion. The presentation will conclude by considering how level-controlled AI-generated texts such as these can be integrated into courses, and the ongoing implications for material development.