Tiny Aya Global
Cohere Labs · Kanada · 2026
Cohere Labs' 3.35-billion-parameter multilingual model covering 70+ languages, Polish included, in an 8K window - the balanced variant of a five-model family, released under a non-commercial licence and behind an access request.
Tiny Aya Global is the balanced member of Tiny Aya, a family of five 3.35-billion-parameter multilingual models published by Cohere Labs on 13 February 2026. The point of the family is coverage rather than size: one small model that holds more than seventy languages at once, including many that larger models treat as an afterthought, and that is small enough to run on hardware a research group or a single developer actually owns. The metadata on the release lists 67 languages explicitly - Polish among them - spanning Europe, South and East Asia, the Middle East and sixteen African languages. Architecturally it is an auto-regressive transformer of Cohere's second-generation design, and the interesting detail is how it spends attention. Three layers out of every four use sliding-window attention with a 4096-token window and rotary positional encoding, handling local context cheaply; the fourth uses global attention with no positional embeddings at all, letting any token reach any other across the whole sequence. The window is 8K tokens in and 8K out - modest by 2026 standards, and the honest trade-off for a model this small carrying this many languages. Global is the instruction-tuned variant: it starts from tiny-aya-base and adds supervised fine-tuning plus preference training, and Cohere describes it as the best balance across languages and regions. Where a deployment is regional, the family offers three siblings tuned for narrower groups - water for European and Asia-Pacific languages, fire for South Asian, earth for West Asian and African. Two restrictions matter more than any benchmark here, and neither is visible in the phrase "open weights". The licence is CC BY-NC 4.0, coupled with Cohere Labs' acceptable use policy: this is a research release, and commercial use is not granted at any scale. The weights are also gated - downloading requires an account and agreeing to share contact information with the publisher. Cohere published its own GGUF quantisations three days after the release, on 16 February 2026, which is what makes local use practical. By 17 August 2026 the safetensors repository had been downloaded 5,912 times and the GGUF build a further 2,755, with 170 likes: the most popular member of the family by a wide margin.
▸Videos
No videos yet.
▸Reviews
No reviews yet. Be the first!