InkubaLM

InkubaLM — named after the dung beetle — is the first multilingual large language model designed for African languages, launched in August 2024 by Johannesburg-based Lelapa AI. It covers five languages (Swahili, Yoruba, isiXhosa, Hausa, isiZulu) spoken by roughly 364 million people, and was trained on 1.9 billion tokens assembled from open datasets on Hugging Face, GitHub, and Zenodo. The model ships open-source alongside two companion datasets, Inkuba-Mono (monolingual pre-training data) and Inkuba-Instruct (instruction tuning for translation, sentiment analysis, NER, QA, and news classification). Training compute was supported by Microsoft's AI for Good Lab.\n\nThe honest picture: InkubaLM is deliberately compact — built for communities without heavy compute, not to win general benchmarks. Coverage is five languages rather than the continent's full linguistic diversity, and low-resource translation quality will lag far behind high-resource languages. But the World Bank has since cited it as an example of specialized local-language models outperforming much larger ones on their target languages.