Skip to content
SW

smallm-wolof

Small Wolof model, MIT

Official site consultedAI: models

Last verified: 1 October 2026

Very small Wolof model of 16.5 million parameters, mixture-of-experts (4 experts, top-2 routing), with an 8,192-token BPE tokenizer built for Wolof. Trained on about 9.7 million tokens (Wolof Wikipedia, FineWeb2, MADLAD-400). The card states it is not suitable for production. MIT license.

Services
  • Wolof text generation
Access requirements

Open weights on Hugging Face, MIT license; trust_remote_code=True required.

Sources