Enterprise AI company Cohere has introduced its latest innovation, the Tiny Aya family of multilingual models, at the India AI Summit. These models, known for their open-weight design, support over 70 languages and can run on everyday devices without an internet connection.
The base model features 3.35 billion parameters and includes specialized versions like TinyAya-Global for broad language support, as well as regional variants such as TinyAya-Earth, TinyAya-Fire, and TinyAya-Water, catering to specific language groups across different regions.
Cohere emphasized that the models were meticulously trained to incorporate linguistic nuances and cultural context, enhancing their natural feel and reliability for diverse user communities. This development is particularly beneficial for developers and researchers creating applications for non-English speaking audiences, especially in regions like India.
These models, optimized for offline usage, can power on-device translation with minimal computing resources, making them highly efficient compared to similar solutions in the market. Cohere’s strategic use of Nvidia’s H100 GPUs in training these models underscores their commitment to enabling robust language processing capabilities on modest hardware configurations.
Source: TechCrunch