Tiny Aya Fire

Tiny Aya Fire is the member of the Tiny Aya family that Cohere Labs recommends for South Asian languages. It exists because of a limit the publisher chose to admit rather than paper over: a 3.35-billion-parameter model cannot carry seventy languages equally well, so instead of one compromise the family offers a balanced variant and three regionally sharpened ones. Fire is the sharpening towards the languages of the Indian subcontinent - among them Hindi, Bengali, Marathi, Gujarati, Punjabi, Tamil, Telugu, Nepali and Urdu, a group that between them serves well over a billion speakers and is chronically thin in models this size. The machinery is the family's: an auto-regressive transformer of Cohere's second-generation design, sliding-window attention over 4096 tokens in three layers out of four with rotary encoding, global attention without positional embeddings in the fourth, and an 8K-token window in and out. All four instruction-tuned variants start from the same tiny-aya-base trunk, so what separates them is the alignment stage rather than scale or architecture. The licence is the family's too, and it is the part a prospective user should read first: CC BY-NC 4.0 with Cohere Labs' acceptable use policy, which permits research use and does not grant commercial use at any scale - a real constraint for anyone hoping to build a South Asian language service on a model that is otherwise unusually well suited to it. The weights are gated behind an access request; official GGUF quantisations arrived on 16 February 2026. By 17 August 2026 the repository had 968 downloads and 28 likes, with a further 193 downloads of the GGUF build.

▸Videos

No videos yet.

▸Reviews