Cohere Introduces New Multilingual AI Models at India AI Summit

Enterprise AI firm Cohere unveiled a new suite of multilingual models, called Tiny Aya, during the ongoing India AI Summit. These open-weight models, available for public use and modification, support over 70 languages and can function on standard devices like laptops without internet access.

Developed by Cohere Labs, the models cater to South Asian languages including Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi.

The primary model consists of 3.35 billion parameters, signifying its depth and versatility. Additionally, Cohere has introduced TinyAya-Global, designed to effectively respond to user commands across various applications. Regional versions have been added: TinyAya-Earth for African languages, TinyAya-Fire for South Asian languages, and TinyAya-Water for languages in Asia Pacific, West Asia, and Europe.

According to the company, “This strategy fosters robust linguistic and cultural understanding, enabling systems that resonate more authentically with the communities they serve. Furthermore, all Tiny Aya models maintain expansive multilingual capabilities, serving as adaptable foundations for continued research and development.”

Cohere asserts that these models, which were designed using a cluster of 64 Nvidia H100 GPUs with relatively basic computing requirements, are particularly suited for app developers targeting native-language audiences. With the ability to operate directly on devices, they also enable offline translation solutions. The software’s architecture is intended for lower power consumption compared to many similar models.

In linguistically rich nations such as India, this offline functionality paves the way for a myriad of applications, eliminating the reliance on continuous internet connectivity.

The models can be accessed via HuggingFace, a leading platform for AI model sharing and testing, as well as through the Cohere Platform. Developers are encouraged to download them from HuggingFace, Kaggle, and Ollama for local use. The company plans to release training and evaluation datasets on HuggingFace alongside a detailed technical report outlining the training approach.

Aidan Gomez, the startup’s CEO, expressed intentions for a public offering in the near future. As reported by CNBC, the company concluded 2025 on an optimistic note, achieving $240 million in annual recurring revenue, with a 50% increase quarter-over-quarter throughout the year.