Enterprise AI company Cohere has launched a new multilingual model suite on the sidelines of the ongoing India AI Summit. This model, called Tiny Aya, is open-ended. This means that the underlying code is publicly available for anyone to use and modify. It supports over 70 languages and can run on everyday devices like laptops, even without an internet connection.
The model, launched by the company’s research arm Cohere Labs, supports South Asian languages such as Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu and Marathi.
The base model contains 3.35 billion parameters that measure size and complexity. Cohere also released TinyAya-Global, a fine-tuned version to better follow user commands for apps that require broad language support. Regional variations complete the range. TinyAya-Earth for African languages; TinyAya-Fire for South Asian languages; TinyAya-Water for Asia Pacific, West Asia and Europe.
“This approach allows each model to develop a stronger linguistic foundation and cultural nuances, creating a more natural and reliable system for the communities it will serve. At the same time, all Tiny Aya models maintain extensive multilingual coverage, making them a flexible starting point for further adaptation and research,” the company said in a statement.
Cohere notes that these models, trained on a single cluster of 64 H100 GPUs (a type of high-performance chip from Nvidia) using relatively modest compute sources, are ideal for researchers and developers building apps for native-language audiences. This model can run directly on the device, allowing developers to use it to power offline translation. The company has built the base software to be suitable for on-device use, noting that it requires less computing power than most similar models.
In a linguistically diverse country like India, these kinds of offline-friendly features can open up a variety of applications and use cases without constant Internet access.
The model is available on the HuggingFace and Cohere platforms, popular platforms for sharing and testing AI models. Developers can download it from HuggingFace, Kaggle, and Ollama for local deployment. The company is also making its training and evaluation datasets for HuggingFace public and plans to release a technical report detailing its training methodology.
Tech Crunch Event
Boston, Massachusetts
|
June 23, 2026
The startup’s CEO Aidan Gomez said last year that the company planned to go public “soon.” According to CNBC, the company ended 2025 on a strong note, posting $240 million in annual recurring revenue, a 50% quarter-on-quarter growth for the year.
