Open-sourcing our first models on Hugging Face
The first French-pivot models for Cameroonian languages are public, open, and downloadable. Here's what shipped and why we gave it away.
Today the first of our French-pivot translation models are live and open on Hugging Face under flagship-ai. Anyone can download them, run them locally, inspect them, and build on them.
What shipped
Compact MarianMT models in both directions through the French hub — X → français and français → X — trained on the corpus we compiled (scripture as the aligned backbone, plus the other written sources we gathered) and unified under the AGLC alphabet. The collection grows as we clean and validate each language; the architecture means adding a language is just adding two more small models.
Why open
A preservation project that locks its outputs behind an API has missed its own point. These languages belong to their communities, not to us. Open models mean:
- A researcher in Yaoundé can run them offline, with no bill and no gatekeeper.
- Mistakes are auditable — anyone can see where a model is weak and help fix it.
- If we disappear tomorrow, the work survives.
They are first models, trained on narrow data, with all the limits we've written about. They are also a real, running starting point that did not exist before — and now belongs to everyone.