Enterprise AI company Cohere has unveiled a new suite of multilingual models, named Tiny Aya, at the India AI Summit. These innovative models are designed to be open-weight, meaning their foundational code is publicly accessible for developers and researchers to utilize and adapt. A significant feature of the Tiny Aya family is its extensive language support, covering over 70 languages, making it a versatile tool for global applications. Crucially, these models boast the capability to operate efficiently on common devices like laptops without requiring an internet connection, thus enabling robust offline functionality.

The development, spearheaded by Cohere’s research division, Cohere Labs, places a particular emphasis on supporting South Asian languages. This includes critical languages such as Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi, addressing a substantial linguistic diversity. The core of this family is the base model, which comprises 3.35 billion parameters, indicating its computational capacity and complexity in processing language tasks.

Beyond the base model, Cohere has introduced specialized variants to cater to diverse linguistic requirements. TinyAya-Global is a version specifically fine-tuned to enhance its ability to follow user commands, making it ideal for applications demanding broad language support. Additionally, regional variants have been developed: TinyAya-Earth targets African languages; TinyAya-Fire is dedicated to South Asian languages; and TinyAya-Water focuses on languages across Asia Pacific, West Asia, and Europe. Cohere emphasized that this tailored approach allows each model to achieve deeper linguistic grounding and cultural nuance, resulting in systems that are more intuitive and reliable for their intended communities, while maintaining extensive multilingual coverage for flexibility in research and adaptation.

Cohere highlighted that these models were trained using relatively modest computing resources, specifically on a single cluster of 64 H100 GPUs from Nvidia. This efficient training methodology makes them particularly suitable for researchers and developers who aim to build applications for audiences speaking native languages. The on-device capability is a game-changer, allowing developers to implement offline translation features, which is especially impactful in linguistically diverse nations like India, where consistent internet access might not always be available. This self-contained operational capacity significantly broadens the potential applications and use cases for these models.

The Tiny Aya models are readily available across popular platforms for AI model sharing and deployment. Developers can access them on HuggingFace, Kaggle, and Ollama for local deployment. Furthermore, Cohere is making the associated training and evaluation datasets available on HuggingFace and has announced plans to release a comprehensive technical report that will detail its training methodology, fostering transparency and further research.

In terms of corporate developments, Cohere's CEO, Aidan Gomez, indicated last year that the company intends to go public in the near future. Financially, Cohere concluded 2025 on a strong note, achieving an impressive $240 million in annual recurring revenue, coupled with a robust 50% quarter-over-quarter growth throughout the year, underscoring its significant presence and growth trajectory in the enterprise AI market.