A new arXiv preprint introduces Pearl-Chat, an autoregressive language model built specifically for Luganda, a Bantu language spoken in Uganda. The authors state that most foundation models target high-resource Indo-European languages, which imposes both computational and morphological penalties on under-represented African languages. They argue that generic multilingual tokenizers are a key part of the problem, as they are not designed for Luganda's agglutinative morphology.

The model is implemented from first principles in JAX and Flax NNX, a choice that suggests a focus on low-level control over the training and inference stack. The abstract frames this as a direct response to the one-size-fits-all approach of existing multilingual models, but it does not yet include benchmark numbers or comparisons. As a single-source preprint, the claims here are the authors' own, and the practical performance of Pearl-Chat remains to be seen.