Building the first local Lithuanian AI model

LEMMIS (Lithuanian Mixture of Experts) is a prototype language model developed from scratch for Lithuanian. The accompanying paper presents the project’s motivation, data collection strategy, tokenizer design, mixture-of-experts architecture, training procedure, and evaluation results. This Zenodo entry provides the paper itself as a standalone research document. Code and related implementation files are maintained separately on Hugging Face. The paper is intended as a technical report on a research prototype. It is not a claim of a production-ready or commercially deployed system.