Launching Aya: New Multilingual AI from Cohere Labs

1 views Source
Launching Aya: New Multilingual AI from Cohere Labs

Aya is a global open-science initiative from Cohere Labs that brings together researchers to advance multilingual AI, bridging gaps between people and cultures worldwide. Recently introduced was Tiny Aya, a compact multilingual AI model that runs locally on any device. This 3.35 billion parameter model supports over 70 languages and offers specialized variants for different regions, delivering strong performance without cloud dependency.

Tiny Aya includes several variants optimized for balanced multilingual performance. For instance, Tiny Aya Earth is designed for languages across Africa and West Asia, Tiny Aya Fire targets South Asian languages, while Tiny Aya Water is tailored for the Asia-Pacific and European regions. Each of these models is developed with real-world scenarios in mind, making them particularly useful.

Additionally, Aya Vision is a research model that advances multilingual multimodal AI through innovative synthetic data generation and model merging. It achieves state-of-the-art performance across 23 languages, surpassing larger models while efficiently addressing data scarcity and catastrophic forgetting by reducing computational overhead by 40% through optimized training techniques.

Aya Expanse redefines multilingual AI by mastering 101 languages through innovative instruction tuning and cross-lingual transfer techniques. By combining a curated open-source dataset with compute-efficient pretraining, this model achieves unparalleled performance across both high- and low-resource languages while reducing infrastructure costs by 30%, setting a new benchmark for scalable, inclusive language modeling.

Aya began as the largest open science collaboration in ML, uniting a community of researchers from around the world to create a powerful foundation for future innovation. This initial effort laid the groundwork for subsequent research initiatives and the development of additional models, pushing the boundaries of what is possible and expanding the world that AI can see.

Related articles