Lelapa AI: African Language Models and Speech Tools

Language AI for constrained settings

Lelapa AI develops language systems with an emphasis on limited computing resources, fragmented data and African linguistic contexts. Its Vulavula products provide speech transcription and text translation, with live and batch workflows for multilingual conversations. Those services are distinct from its downloadable research models. [1][2]

The July 2024 InkubaLM release introduced a 0.4B-parameter base language model and the Inkuba-Mono and Inkuba-Instruct datasets. Training covered isiZulu, Yoruba, Hausa, Swahili and isiXhosa, with English and French also included. The announcement linked the artifacts and explicitly withheld the instruction test set at that stage; the later paper submission is not used as the release date. [3][5]

The official InkubaLM checkpoint lists CC BY-NC 4.0, and its current Hugging Face repository requires agreement and account-information sharing before file access. The noncommercial model terms should not be conflated with the terms for Vulavula’s commercial APIs or with anonymous unrestricted downloads. [4]