At the India AI Summit, enterprise AI firm Cohere introduced a new series of multilingual models. Named Tiny Aya, these open-weight models—meaning their core code is accessible for public use and modification—support more than 70 languages and can operate on common devices like laptops without needing an internet connection. Developed by Cohere Labs, the company’s research division, the models include support for South Asian languages such as Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. The base model features 3.35 billion parameters, indicating its scale and sophistication.

Cohere has also released TinyAya-Global, a version refined to better adhere to user instructions, designed for applications needing wide language coverage. The family is completed with regional variants: TinyAya-Earth for African languages, TinyAya-Fire for South Asian languages, and TinyAya-Water for Asia Pacific, West Asia, and Europe.

In a statement, the company explained, “This approach allows each model to develop stronger linguistic grounding and cultural nuance, creating systems that feel more natural and reliable for the communities they are meant to serve. At the same time, all Tiny Aya models retain broad multilingual coverage, making them flexible starting points for further adaptation and research.” Cohere highlighted that these models were trained on a single cluster of 64 H100 GPUs, a high-performance chip from Nvidia, using relatively modest computational resources, making them well-suited for researchers and developers creating apps for native language speakers.

The models are designed to run directly on devices, enabling developers to implement offline translation features. Cohere emphasized that its underlying software is optimized for on-device use, requiring less computing power than many similar models.

In linguistically diverse regions like India, this offline capability can unlock a wide range of applications without relying on constant internet access. The models are accessible on HuggingFace, a popular platform for sharing and testing AI models, as well as on the Cohere Platform. Developers can download them from HuggingFace, Kaggle, and Ollama for local deployment. Cohere is also releasing training and evaluation datasets on HuggingFace and intends to publish a technical report detailing its training methodology.

Last year, Cohere’s CEO Aidan Gomez stated that the company plans to go public “soon.” According to CNBC, Cohere finished 2025 strongly, achieving $240 million in annual recurring revenue with 50% quarter-over-quarter growth throughout the year.