Enterprise AI firm Cohere has announced the launch of a new suite of multilingual models, known as Tiny Aya, at the ongoing India AI Summit. These models are open-weight, meaning their foundational code is accessible for public use and modification. They support more than 70 languages and can efficiently operate on everyday devices like laptops without requiring internet access, making them valuable tools for a variety of applications.
Developed by Cohere Labs, the Tiny Aya models specifically cater to South Asian languages, including Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. The base model boasts 3.35 billion parameters, reflecting its complexity and capability. In addition to the main model, Cohere has introduced TinyAya-Global, which is optimized for improved response to user commands and broad language compatibility. The lineup includes regional adaptations: TinyAya-Earth for African languages, TinyAya-Fire for South Asian languages, and TinyAya-Water for audiences in Asia Pacific, West Asia, and Europe.
Cohere emphasizes that this localized approach enhances linguistic accuracy and cultural relevance, ensuring the models feel natural for the communities they serve. Notably, Tiny Aya models retain extensive multilingual capabilities, offering a solid foundation for further customization and research.
Trained on a robust cluster of 64 Nvidia H100 GPUs, these models are designed with on-device use in mind, requiring less computational power than many of their competitors. This makes them particularly suitable for researchers and developers focusing on applications for native language speakers. The ability to function offline is especially advantageous in linguistically diverse nations like India, as it supports a wide array of applications without the dependency on constant internet connectivity.
Developers can access the Tiny Aya models on platforms such as Hugging Face and Cohere’s own platform, with options for local deployment available through Kaggle and Ollama. Additionally, the company plans to release training and evaluation datasets on Hugging Face, along with a forthcoming technical report detailing the training methodologies employed.
Cohere’s CEO, Aidan Gomez, has previously indicated that the company aims to go public soon. According to recent reports from CNBC, the company saw robust performance, wrapping up 2025 with $240 million in annual recurring revenue—a 50% growth from previous quarters.
