Cohere Labs launches Tiny Aya multilingual AI model

Cohere Labs has launched Tiny Aya, a compact multilingual model designed to run comfortably on consumer hardware and mobile phones. The base model weighs in at roughly 3.35 billion parameters and supports more than 70 languages.

The model targets translation and general-purpose chat across multiple regions. There’s also a fine‑tuned variant called Tiny Aya Global, narrowing in on 67 key languages spanning five global regions.

In addition, benchmarking shows it outperforming similarly sized models on multilingual generation quality. Efficiency is its key selling point with plots of quality versus parameter count placing Tiny Aya toward the ‘upper-left’ sweet spot. For reference, this means it combines strong output quality with a small footprint.

The model is already available as open weights on Hugging Face, where users can try live demos for prompts. Languages available include Spanish, Arabic, Chinese, Basque-to-French translation, and more.

There’s noticeably fast responses thanks to its compact size with the Tiny Aya Global checkpoint being around 6.7 GB. This fits neatly on most mid-range consumer GPUs and opens up realistic on-device multilingual AI deployments.