Feb 18, 2026 20:30:00

Enterprise AI developer Cohere launches multilingual AI 'Tiny Aya'

Cohere Labs, the research division of Cohere, which develops enterprise AI for businesses, has released a multilingual open-weight model called ' Tiny Aya .' Cohere Labs developed the model with the goal of 'maintaining reliability across many languages while remaining efficient enough for local use.' The model is said to be competitive with existing multilingual models in tasks such as verification, translation, and mathematical inference for low-resource languages.

Introducing ✨Tiny Aya✨, a family of massively multilingual small language models built to run where people actually are.

Tiny Aya delivers strong multilingual performance in 70+ global languages in a 3.35B parameter model, efficient enough to run locally, even on a phone.

pic.twitter.com/TV8ESmFCrp

— Cohere Labs (@Cohere_Labs) February 17, 2026

Cohere Labs Launches Tiny Aya, Making Multilingual AI Accessible
https://cohere.com/blog/cohere-labs-tiny-aya

Cohere launches a family of open multilingual models | TechCrunch
https://techcrunch.com/2026/02/17/cohere-launches-a-family-of-open-multilingual-models/

The benchmark results for multilingual translation tasks performed on four lightweight models—Tiny Aya, Google Gemma 3-4b , Mistral AI Ministral 3-3b , and Qwen 3-4b— are shown below. As Cohere aims, Tiny Aya outperforms the others in West Asian and African languages.

The larger the number of pages on

Common Crawl , which provides web archives and datasets spanning several petabytes, the greater the interlingual performance of the multilingual LLM. The yellow-green line in the graph below shows the number of pages in Common Crawl, and Tiny Aya demonstrates stable performance even for languages less commonly used on the web, which have fewer pages on Common Crawl.

Below is a graph showing the model size (horizontal axis) of open-ended generation AI and the quality (vertical axis) averaged over 66 languages. In terms of quality, Gemma 3-27b is the best, but the model size is very large, while Tiny Aya shows high quality while keeping the model size small.

Bar chart showing tokenization efficiency by language. Tiny Aya achieves the most efficient tokenization for most languages.

Tiny Aya is prioritizing post-training in 67 languages across five regions around the world.

There are four specialized models: 'Tiny Aya Fire,' 'Tiny Aya Earth,' 'Tiny Aya Water,' and 'Tiny Aya Global,' each with its own unique strengths in each region.

Tiny Aya is available as an open weight model and is available to try out immediately on the Hugging Face Space and Cohere platforms.

Tiny Aya - a Hugging Face Space by CohereLabs
https://huggingface.co/spaces/CohereLabs/tiny-aya?ref=cohere-ai.ghost.io