Enterprise AI startup Cohere has unveiled a new suite of compact multilingual language models during the India AI Summit, aiming to make advanced AI more accessible across diverse linguistic regions. The new lineup, called Tiny Aya, features open-weight models allowing developers and researchers to freely access, modify, and build upon the underlying code. Designed for efficiency, the models support more than 70 languages and can operate directly on devices such as laptops without requiring continuous internet connectivity.
Developed by Cohere Labs, the base model includes 3.35 billion parameters and covers several South Asian languages, including Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. Alongside the core version, the company introduced TinyAya-Global, optimized for stronger instruction-following capabilities across broader language use cases.
The family also includes region-focused variants: TinyAya-Earth for African languages, TinyAya-Fire tailored to South Asia, and TinyAya-Water covering Asia Pacific, West Asia, and parts of Europe. According to the company, this regional approach strengthens cultural and linguistic depth while maintaining wide multilingual adaptability for research and customization.
Cohere stated that the models were trained using a relatively compact infrastructure powered by a single cluster of 64 H100 GPUs from Nvidia making them efficient compared to many large-scale AI systems. Their optimized architecture enables offline deployment, which could be especially valuable in countries like India where internet access may not always be consistent.
The Tiny Aya models are available through platforms such as Hugging Face, Kaggle, and Ollama, as well as via the Cohere Platform. The company is also releasing supporting training and evaluation datasets and plans to publish a detailed technical report outlining its development process.
Founded by CEO Aidan Gomez, Cohere has been expanding its enterprise AI footprint. Reports indicate the company closed 2025 with approximately $240 million in annual recurring revenue, reflecting strong quarterly growth as it positions itself for a potential public listing in the near future.
