Tiny Aya in Practice: What You Can Build with Open Multilingual AI

Cohere Labs' Tiny Aya is a lightweight open-weight model that supports over 70 languages and can run locally even on smartphones. Through the recently conducted 'Expedition Tiny Aya' program, a variety of research achievements created by developers around the world using this model have been revealed.

The main research themes are summarized into four areas.

• Education and Learning: Tiny Aya Math Edition, which transparently shows the mathematical reasoning process, and Kids Companion, a multilingual voice assistant for children that works in offline environments, were developed.
• Multilingual Safety: Researchers discovered that safety behavior changes in situations involving code-mixed prompts or inconsistencies between languages, presenting the need for a new evaluation framework.
• Accessibility and Local Deployment: Compression techniques and synthetic data pipelines were developed for use in mobile and offline environments, improving accessibility for underserved languages.
• Language Understanding: It was demonstrated that Tiny Aya possesses deep semantic cross-lingual transfer capabilities, such as word sense disambiguation, going beyond simple translation.