Enterprise AI company Cohere unveiled a new family of multilingual models on the sidelines of the ongoing India Artificial Intelligence Summit. The models, called Tiny Aya, are open-source — meaning their underlying code is publicly available for use and modification — support more than 70 languages, and can run on everyday devices like laptops without requiring an internet connection.
Released by the company’s research arm Cohere Labs, the model supports South Asian languages such as Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu and Marathi.
The base model contains 3.35 billion parameters — a measure of its size and complexity. Cohere also released TinyAya-Global, a version tailored to better follow user commands, for applications that require broad language support. Local variants complete the family: TinyAya-Earth for African languages. TinyAya-Fire for South Asian languages. and TinyAya-Water for Asia Pacific, West Asia and Europe.
“This approach allows each model to develop a stronger linguistic base and cultural nuance, creating systems that feel more natural and authentic to the communities they are intended to serve. At the same time, all Tiny Aya models maintain broad multilingual coverage, making them flexible starting points for further customization and research,” the company said in a statement.
Cohere noted that these models, which were trained on a single cluster of 64 H100 GPUs (a type of high-powered chip from Nvidia) using relatively modest computing resources, are ideal for researchers and developers building apps for native-speaking audiences. Models can run directly on devices, so developers can use them to enable offline translation. The company noted that it built its underlying software to suit on-device use, requiring less computing power than most comparable models.
In linguistically diverse countries like India, this kind of offline-friendly capability can open up a wide range of applications and use cases without the need for constant Internet access.
The models are available on HuggingFace, the popular platform for sharing and testing artificial intelligence models, and on the Cohere platform. Developers can download them to HuggingFace, Kaggle and Ollama for local development. The company is also publishing HuggingFace training and evaluation datasets and plans to publish a white paper detailing its training methodology.
Techcrunch event
Boston, MA
|
June 23, 2026
The startup’s CEO, Aidan Gomez, said last year that the company plans to go public “soon.” According to CNBCthe company closed out 2025 on a high note, posting annual recurring revenue of $240 million, up 50% quarter-over-quarter for the full year.
