Most AI models treat African languages as an afterthought. MORENA, released last month by South African startup Vambo AI, was built around them. It is a 1.5-billion-parameter open-weight model trained from scratch for ChiShona, Kiswahili, Hausa, Yoruba, Igbo, isiZulu, isiXhosa, Kinyarwanda, Setswana, Afrikaans, isiNdebele and Nigerian Pidgin, plus English and French, according to CIO Africa.
Why from scratch matters
Most African-language projects take an existing model such as Llama and keep training it, which means keeping a vocabulary built for English and code, TechRadar explains. Vambo AI designed the vocabulary first. MORENA encodes African text with 1.39 times fewer tokens than Google's Gemma 3 and 1.53 times fewer than Meta's Llama 3.2, and fewer tokens means lower cost every time the model is used.
The claims
The company says the base model scores 1.408 bits per byte on African-language modelling, the best of 26 models it tested, ahead of Lugha-Llama-8B at 1.423, a model about five times its size. Co-founder Isheanesu Misi says it outperforms a Google model about eight times larger; the company has not named that model. These are the developer's own results, CIO Africa notes.
Training used 251.7 billion tokens and about 22,000 GPU hours worth roughly $40,000, with support from UNDP, AIHub4SD and Italy's CINECA, according to TechRadar. Smaller 0.5-billion and 0.2-billion versions are also available, the smallest aimed at keyboards and speech tools.
