South Africa's Vambo AI Launches Model for African Languages

South African startup Vambo AI has released MORENA, an open-weight 1.5-billion-parameter language model. Founded in April 2023 by Chido Dzinotyiwei and Isheanesu Misi, Vambo AI aims to build artificial intelligence that understands the languages of people globally. The MORENA model was trained from scratch for 12 African languages, as well as English and French, marking a significant step in localized AI development.

The Vambo AI platform is designed to serve both individual users and businesses. Individual users can leverage its capabilities to write, search, translate, and transcribe content across various languages. For businesses and innovators, the startup provides an API and developer tools, enabling them to build tailored solutions for diverse geographies and localize existing solutions that have successfully addressed problems in other regions.

MORENA specifically understands ChiShona, Kiswahili, Hausa, Yorùbá, Igbo, isiZulu, isiXhosa, Kinyarwanda, Setswana, Afrikaans, isiNdebele, and Nigerian Pidgin. It also includes support for English and French. Isheanesu Misi, co-founder of Vambo AI, stated that MORENA is the largest open African AI model trained from scratch. He noted that it outperforms a Google model eight times its size on African text modeling and is over three times the size of the second largest African language model.

The model's efficient tokenization of African languages makes it both cheaper to run and adapt within African contexts. This efficiency is expected to significantly benefit startups in countries like Nigeria and researchers in Kenya, as they will spend less to develop and deploy their own AI models in real-world settings. Misi emphasized that MORENA proves African initiatives from scratch do not necessarily require billion-dollar budgets and provides other researchers with a new point of reference.

Vambo AI recommends fine-tuning as the optimal method for utilizing MORENA. The release of this new model significantly expands the platform's existing capabilities, which previously understood 11 languages, including Arabic, KiSwahili, IsiZulu, and French. This development aims to foster local innovation by providing accessible and cost-effective language technology solutions across the continent.