An AI Learns 12 African Languages. When Will Bulu and Duala Follow?

MORENA is trained around African languages. Could researchers in Cameroon develop similar tools for the country's many local languages?

By AgenStudio15 October 2026Updated 29 October 20266 min read

Share

LinkedIn Facebook WhatsApp X

Quick read

What to take away

Three anchors to enter the analysis without losing the thread.

01

Why the language list matters

MORENA focuses on Swahili, Hausa, Yoruba, Igbo, Shona, Kinyarwanda, Zulu, Xhosa, Setswana, Afrikaans, Southern Ndebele and Nigerian Pidgin.

02

What the benchmark can and cannot prove

The model's published technical card compares performance using bits per byte: a measurement of how efficiently a model predicts test text. On that specific benchmark, lower is…

03

What could Cameroon do?

A first project need not support every language or build a massive chatbot. A university and a community of consenting speakers could document one well-defined task, create a…

A South African team built an AI model around twelve African languages. In Cameroon, where hundreds of languages coexist, that raises a bigger question: can our own communities help create the next model?

Vambo AI's MORENA was trained from scratch, with a model vocabulary designed for the languages it targets. Its base version has approximately 1.5 billion parameters — numerical settings learned during training.

Why the language list matters

MORENA focuses on Swahili, Hausa, Yoruba, Igbo, Shona, Kinyarwanda, Zulu, Xhosa, Setswana, Afrikaans, Southern Ndebele and Nigerian Pidgin.

Hausa is also spoken in Cameroon. But Ewondo, Bulu and Duala are not among the twelve listed languages. Nor should Nigerian Pidgin automatically be treated as equivalent to all forms of Cameroonian Pidgin.

A language model needs much more than a dictionary: reliable texts, speakers who check examples, thoughtful treatment of dialects and consent to use recordings.

What the benchmark can and cannot prove

The model's published technical card compares performance using bits per byte: a measurement of how efficiently a model predicts test text. On that specific benchmark, lower is better.

The published results include 1.408 for MORENA 1.5B base, 1.423 for Lugha-Llama 8B and 2.159 for Gemma 3 12B Instruct.

MORENA: African language text prediction

Lower bits per byte indicates better performance on the authors' specific text dataset.

Sort by

Prediction · bits per byte

Visible categories : 3 / 3

Vambo AI — fiche MORENA

That is a text-prediction comparison, not a universal ranking of intelligence or a guarantee of translation accuracy. The model weights use the Apache 2.0 licence, but the full training corpus is not released alongside them.

What could Cameroon do?

A first project need not support every language or build a massive chatbot. A university and a community of consenting speakers could document one well-defined task, create a small evaluation set and measure whether a model handles it faithfully.

Applications in schooling or routine administration would need human review; errors in medicine or legal advice can be dangerous.

The real opportunity is to develop African-language technology with the people whose languages it aims to represent, including how their data is collected and used.

Africa is showing that local languages can influence the design of AI. Cameroon can help determine what comes next.