Enterprise AI company Cohere has unveiled a new family of multilingual AI models, expanding global access to artificial intelligence tools that support diverse languages and run directly on everyday devices. The models, introduced during the India AI Summit, aim to improve language accessibility while enabling developers to build AI-powered applications without constant internet connectivity.

Cohere Introduces Multilingual AI Models With Offline Capabilities

Cohere’s new multilingual AI models, called Tiny Aya, are open-weight systems, meaning their core code is publicly accessible for developers and researchers to modify and adapt. The models support more than 70 languages and can operate locally on devices such as laptops without requiring internet access.

Developed by Cohere Labs, the models include support for several South Asian languages, including Hindi, Bengali, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. This approach is designed to make AI tools more accessible to diverse global communities.

The base version of the model contains 3.35 billion parameters, reflecting its scale and complexity. Cohere also introduced TinyAya-Global, a version fine-tuned to better understand and follow user instructions for applications requiring broad multilingual functionality.

Regional Variants Expand Language Coverage

Cohere has released several regional versions of its multilingual AI models to strengthen linguistic understanding and cultural accuracy across different regions.

The lineup includes TinyAya-Earth for African languages, TinyAya-Fire for South Asian languages, and TinyAya-Water designed for Asia Pacific, West Asia, and Europe. According to the company, this regional approach allows each model to develop stronger cultural and linguistic context while maintaining broad multilingual capabilities.

Cohere said the strategy helps create AI systems that feel more natural and reliable for local users while remaining flexible for research and application development.

Built for Researchers and Developers

The company emphasized that the multilingual AI models were trained using relatively modest computing resources, relying on a single cluster of 64 Nvidia H100 GPUs. The models are optimized for on-device deployment and require less computational power compared to many similar systems.

Because they can run offline, the models enable developers to build tools such as local translation systems without continuous internet access. This capability is particularly valuable in linguistically diverse regions where connectivity may be limited.

Cohere noted that such offline-friendly technology could unlock new applications in countries like India, supporting language inclusion and broader digital adoption.

Availability and Future Plans

Cohere has made the multilingual AI models available through platforms including HuggingFace and the Cohere Platform. Developers can also access them via Kaggle and Ollama for local deployment.

The company is releasing training and evaluation datasets to support research and plans to publish a technical report detailing the model development process.

Cohere continues to expand its enterprise AI footprint. CEO Aidan Gomez previously stated that the company plans to go public soon, and reports indicate the firm ended 2025 with $240 million in annual recurring revenue and strong quarterly growth.