Tiny Aya Earth

Tiny Aya Earth is the member of the Tiny Aya family that Cohere Labs recommends for West Asian and African languages. Of the four instruction-tuned variants this is the one covering the ground that mainstream models cover worst: alongside Arabic, Persian, Hebrew and Turkish, the release metadata lists sixteen African languages, including Amharic, Hausa, Igbo, Malagasy, Shona, Swahili, Wolof, Xhosa, Yoruba and Zulu. For most of these there is no comparable first-party option at three billion parameters, which is the practical reason this profile exists in the catalogue. The engineering is shared across the family: an auto-regressive transformer of Cohere's second-generation design with sliding-window attention over 4096 tokens in three layers out of four, rotary positional encoding in those layers, global attention without positional embeddings in the fourth, and 8K tokens of context in and out. All the regional variants and the balanced Global variant are instruction-tuned descendants of the same tiny-aya-base trunk, so the difference between them lies entirely in the alignment stage. The access terms are the same as the rest of the family and deserve emphasis in this particular case: CC BY-NC 4.0 with Cohere Labs' acceptable use policy permits research use and grants no commercial use at any scale, so an organisation wanting to serve, say, Swahili or Amharic speakers commercially cannot build on these weights, however well suited they are. The files are gated behind an access request, and official GGUF quantisations followed on 16 February 2026. By 17 August 2026 the repository had 1,483 downloads and 25 likes, plus 137 downloads of the GGUF build - the most downloaded of the three regional models.

▸Videos

No videos yet.

▸Reviews