African Tongues and Languages at Scale (ATLAS)
Umoja means unity. Across Africa, more than 2,000 languages are spoken, yet AI understands only a fraction. ATLAS Umoja AI is an open platform bringing together governments, industry, researchers and innovators to build the benchmarks, models and communities needed for African language AI.
2,000+
54
1
Why Atlas umoja ai
The building blocks African AI has been missing
Benchmarks and models for African languages, built with local communities and free for developers to build on.
Benchmarks (opens in new tab)
Evaluation benchmarks built with language communities across the continent — so models can be tested on how well they actually understand African languages, not just approximate them.
Models (opens in new tab)
Multilingual, multimodal models free to download, fine-tune, and deploy. Hosted on Hugging Face so the world can access, evaluate, and improve them.
Community (opens in new tab)
Developers, startups, researchers, and governments build on top — turning shared infrastructure into real tools for real people in real languages.
Our commitment
ATLAS Umoja AI gives every nation and its builders access to the building blocks of AI.
A shared and open framework to accelerate building AI African Tongues and Languages at Scale.
atlas umoja ai
A living map of African language AI
Select a country to explore its open datasets and model. Live countries have open models available on Hugging Face today.
How it works
Each country, in its own identity
ATLAS Umoja AI is a shared framework that each country can use to shape and build its own AI model. One that's based on its local language, dialects and culture. All on a timeline that suits them.
-
Support the next country launch
A national ministry, research institution, or civic organisation steps forward. They anchor the initiative, connect with speaker communities, and own the process.
-
Collect with communities
Language data is gathered through structured programmes — recorded, transcribed, and validated by native speakers who understand the nuance of each language.
-
Train & release openly
A model is trained on the collected data and released on Hugging Face under an open licence. Any developer anywhere can download, fine-tune, and build on it.