A team of linguists and engineers has put out a set of open‑weight language models that understand and generate Wolof, a language spoken by about 23 million people in Senegal. The models range from 0.6 billion to 14 billion parameters and are released under an open licence, so anyone can download and run them locally. The paper says the 8 billion and 14 billion versions perform on par with much larger commercial systems on Wolof tasks while still handling other languages well.
What the paper reports
The authors describe a pipeline that first fine‑tunes base models on Wolof text, then applies instruction tuning (teaching the model how to follow prompts) and preference alignment (optimising outputs to match human judgments). They also merge models to improve performance. Experiments cover 63 public and internal tasks, including translation, question answering, and text generation. Results show the 8 B and 14 B models are especially competitive.
Availability
The models are already being tested in internal customer‑care pilots and the code and weights are hosted at a public URL listed in the paper. No cost is mentioned for downloading, and because the weights are open‑weight, users can run them on their own hardware without paying API fees.
Why it matters
For Wolof speakers and businesses in Senegal, the models provide a free, locally‑hostable alternative to costly API services. This can lower expenses and avoid data‑privacy concerns when building chatbots or translation tools. Independent validation of the claims beyond the authors’ experiments is still pending.
