Vambo AI Releases Open-Weight MORENA Model

South African start-up Vambo AI has released MORENA, an open-weight language model with 1.5 billion parameters that covers 12 African languages, plus English and French, and says it trained it from scratch.

The languages are ChiShona, Kiswahili, Hausa, Yoruba, Igbo, isiZulu, isiXhosa, Kinyarwanda, Setswana, Afrikaans, isiNdebele and Nigerian Pidgin. Open-weight means the trained model parameters are published for others to use and adapt.

Vambo AI co-founder Isheanesu Misi reportedly says MORENA is the largest open African AI model trained from scratch, and that it outperforms a Google model with about eight times as many parameters on African text modelling. The company has not named that model. Remarks attributed to Misi are paraphrased in the coverage reviewed.

TechRadar, citing the developers, reported these results. None has been independently verified:

• The base model scores 1.408 bits per byte on language modelling, the best of 26 models tested. Lugha-Llama-8B, an Africa-adapted Llama 3.1 variant with about five times the parameters, scores 1.423.
• MORENA beats that model in eight of the 12 languages.
• Its tokeniser uses 1.39 times fewer tokens than Google’s Gemma 3 and 1.53 times fewer than Meta’s Llama 3.2 on African text. That can lower the cost of running and adapting the model.
• Pretraining used 251.7 billion tokens and more than 22,000 A100 GPU hours, valued at about $40,000.

TechRadar reported that UNDP, AIHub4SD and CINECA provided computing resources and other support. Vambo AI also reports smaller 0.5-billion and 0.2-billion variants.

Chido Dzinotyiwei and Isheanesu Misi founded Vambo AI in April 2023. Its platform lets individuals write, search, translate and transcribe. Businesses and developers get an API and tools. Vambo AI recommends fine-tuning MORENA as a base model instead of deploying it unadapted.

The coverage reviewed does not state the licence terms, training data sources, download location or funding.