Cohere's new multilingual AI models can run on your laptop
What's the story
Cohere, a leading enterprise AI company, has unveiled a new range of open-weight multilingual models. Named Tiny Aya, these innovative models support over 70 languages and can be run on personal devices like laptops without an internet connection. The launch was announced at the ongoing India AI Impact Summit 2026 by the company's research division, Cohere Labs.
Language support
Tiny Aya supports these South Asian languages
The Tiny Aya models are particularly notable for their support of South Asian languages such as Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu and Marathi. The base model comes with a whopping 3.35 billion parameters. In addition to the base model, Cohere has also introduced TinyAya-Global - a version optimized to follow user commands better for applications requiring extensive language coverage.
Model variants
Regional variants included
The Tiny Aya model family also includes regional variants: 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 stronger linguistic grounding and cultural nuance," the company said in a statement. "At the same time, all Tiny Aya models retain broad multilingual coverage."
Offline capability
Ideal for offline translation and native language apps
The models were trained on a single cluster of 64 NVIDIA's H100 GPUs using modest computing resources. They are ideal for researchers and developers building apps for native language speakers. The models can run directly on devices, making them perfect for powering offline translation. Cohere has designed its underlying software to be lightweight enough to run on-device, needing less computing power than most similar models.
Accessibility
Available on HuggingFace and Cohere platform
The Tiny Aya models are now available on HuggingFace, a popular platform for sharing and testing AI models, and the Cohere Platform. Developers can download them from HuggingFace, Kaggle, and also Ollama for local deployment. The firm is also releasing training and evaluation datasets on HuggingFace, and plans to publish a technical report detailing its training methodology.