Cohere is launching a family of open source multilingual versions

Enterprise AI company Cohere launched a new family of multilingual models on the sidelines of the ongoing India AI Summit. The models, called Tiny Aya, are open source – meaning their code is publicly available for anyone to use and modify – they support more than 70 languages, and can be used on everyday devices such as laptops without the need for an internet connection.

The version, launched by the company’s research team Cohere Labs, supports South Asian languages ​​such as Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi.

The original version has 3.35 billion units – a measure of its size and complexity. Cohere has also implemented TinyAya-Global, a model optimized for user compliance, for applications that require multilingual support. Regional diversity around the family: TinyAya-Earth for African languages; TinyAya-Fire for South Asian languages; and TinyAya-Water in Asia Pacific, West Asia, and Europe.

“This method allows each model to have a strong linguistic base and a cultural system, creating a system that feels more natural and reliable for the communities that are meant to serve. At the same time, all Tiny Aya models keep a wide range of languages ​​​​Learning, making them change the principles from the beginning of other transfer and research,” the company said in its statement.

Cohere noted that these models, which were trained on a single cluster of 64 H100 GPUs (the most powerful chip type by Nvidia) using limited computing resources, are ideal for researchers and developers to create audio programs that speak native languages. The models can run directly on devices, so developers can use them for offline rendering. The company noted that it designed its core software to be compatible with applications on devices, which require less computing power than similar models.

In multilingual countries like India, such easy access to the Internet can open various programs and use cases without the need for constant Internet access.

The models are available on HuggingFace, a popular platform for sharing and testing AI models, and the Cohere Platform. Developers can download from HuggingFace, Kaggle, and Ollama for local deployment. The company is also releasing training and evaluation materials on HuggingFace and plans to release a technical report detailing its training methods.

The company’s CEO, Aidan Gomez, said last year that the company plans to make an announcement “soon”. According to Price CNBCthe company ended 2025 at a high price, delivery $240 million a year over and over again, with a growth rate of 50%. quarter to quarter throughout the year.