Cohere launches a family of open multilingual models

Cohere launched Tiny Aya, a family of open-weight multilingual models, supporting over 70 languages including South Asian languages, designed for offline use on standard devices. The models, ideal for app developers and researchers, were unveiled at the India AI Summit and are accessible on platforms like HuggingFace. This initiative aims to foster better linguistic understanding and cultural nuances in AI applications.

Key Points

• Cohere introduced Tiny Aya models at the India AI Summit.
• Models support over 70 languages, with a focus on South Asian languages.
• They are open-weight, allowing modification and public use.
• The base model has 3.35 billion parameters and can run offline on standard devices.
• There are regional variants: TinyAya-Global, TinyAya-Earth, TinyAya-Fire, and TinyAya-Water.
• Models require less computing power, making them suitable for underserved linguistic communities.
• Cohere plans to release training datasets and a detailed technical report on model training.

Relevance

• The launch aligns with the increasing demand for multilingual AI tools to support diverse populations.
• Models like Tiny Aya respond to the trend of decentralization and affordability in AI technology.
• It reflects the growing emphasis on inclusion and accessibility in AI development.
• Cohere's initiative coincides with the rise of AI applications in regions with limited internet access.

The launch of Tiny Aya by Cohere marks a significant step towards more inclusive AI, catering to diverse linguistic needs while promoting accessibility and local deployment, potentially influencing future AI developments aimed at enhancing cultural relevance.