Cohere, a leader in enterprise AI, has introduced a new series of multilingual models known as Tiny Aya during the India AI Summit. These models, designed with open-weight architecture, allow public access and modification of their code, supporting over 70 languages. Remarkably, they can function on standard devices such as laptops without requiring an internet connection.
Developed by Cohere Labs, the models cater to a variety of South Asian languages including Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. At their core, the models boast 3.35 billion parameters, highlighting their complexity and capacity. Additionally, Cohere has released TinyAya-Global, a refined version tailored to better understand user instructions, ideal for applications demanding extensive language support.
The Tiny Aya model family includes specific regional versions: TinyAya-Earth for African languages, TinyAya-Fire for South Asian languages, and TinyAya-Water for languages across Asia Pacific, West Asia, and Europe. Cohere emphasizes that this tailored approach enhances the linguistic and cultural relevance of each model, making them more intuitive and dependable for their intended communities. Despite these specializations, all Tiny Aya models maintain wide-ranging multilingual capabilities, providing versatile foundations for further innovation and research.
These models were developed on a single cluster of 64 H100 GPUs from Nvidia, demonstrating their efficiency with moderate computing resources. Designed for ease of use on devices, they enable developers to implement offline translation and other applications seamlessly. Cohere's software is optimized for on-device performance, requiring less computational power than many similar models.
In countries with rich linguistic diversity, like India, the ability to operate offline can significantly expand the potential applications and use cases, eliminating the need for constant internet connectivity. The models are accessible on platforms like HuggingFace, Kaggle, and Ollama, offering opportunities for local deployment. Cohere is also sharing training and evaluation datasets on HuggingFace and plans to publish a comprehensive report on their training processes.
The company has ambitious plans for the future, as noted by CEO Aidan Gomez, who indicated intentions to go public soon. Cohere concluded 2025 with a strong financial performance, achieving $240 million in annual recurring revenue and 50% growth quarter-over-quarter, according to CNBC.