Most AI tools still treat African languages as an afterthought. A Johannesburg-based startup, Vambo AI, is trying to change that with MORENA, a new language model built around African speech from the first line of training.
What Vambo has released
MORENA is an open-weight model with 1.5 billion parameters. It was trained from scratch for 12 African languages, plus English and French. The languages are Nigerian Pidgin, Igbo, Yoruba, Hausa, Kiswahili, ChiShona, isiZulu, isiXhosa, Kinyarwanda, Setswana, Afrikaans and isiNdebele. Co-founder Isheanesu Misi describes it as the largest open African AI model trained from scratch. That is the company’s own claim, and no independent ranking has confirmed it.
Open-weight means developers can download the model and adapt it, though that is not the same as fully open-source. Vambo recommends fine-tuning as the best way to use it. Reports also mention smaller 0.5 billion parameter versions for people with limited computing power.
Why training from scratch matters
Many African language projects take an existing model built mainly on English and keep adjusting it. That leaves them stuck with a vocabulary designed for English text. Vambo chose its tokeniser, its data mix and its language list before training began.
A tokeniser breaks text into small pieces the model can process. When a language is split into too many pieces, every request costs more to run. According to the company, MORENA represents African text in fewer pieces than larger rival models, which makes it cheaper to run and adapt.
What the early results show
The developers report that their closest competitor, an Africa-adapted Llama model with eight billion parameters, trailed MORENA in eight of the 12 languages on a standard text-prediction measure. They also say the instruction version translated English into five African languages at a quality level with one dedicated translation system, after seeing just three example translations.
These are developer-reported figures, not independent audits. They are also narrow tests, so they should not be read as proof that MORENA beats every large AI model at everything.
The data problem behind the build
African languages have far less digital text than English, and that shortage shaped the project. Vambo relied partly on synthetic data to fill the gaps, which reportedly limited grammar accuracy in some cases. Training was supported by access to a European supercomputer through a United Nations Development Programme AI hub.
Who stands to benefit
Vambo was founded in April 2023 by Chido Dzinotyiwei and Isheanesu Misi, and has sold language tools to individuals and businesses. A cheap, adaptable base model could help local startups build customer support bots, education tools, translation services and voice products in languages their users actually speak. Nigerian developers working in Yoruba, Igbo, Hausa and Pidgin are among those who could use it.
What to watch next
Proof of MORENA’s value will have to come from outside Vambo. Independent testing, community fine-tuning and real products built on it will show whether it performs as well in practice as it does on benchmarks. Developers should also check output quality in their own language before committing to it, particularly for languages where training data was limited.
For now, MORENA is an encouraging sign that African teams can build language AI on their own terms, with African languages at the centre from day one rather than bolted on later.
