Cohere Labs Launches Tiny Aya to Bring Multilingual AI Directly to Mobile Devices

Cohere Labs has unveiled Tiny Aya, a family of open-weight multilingual models specifically engineered for local deployment on consumer hardware and mobile phones. At just 3.35 billion parameters, these models achieve state-of-the-art translation and reasoning performance across more than 70 languages. This release aims to dismantle barriers to AI access in regions where high-speed internet or cloud infrastructure remains limited.

The collection includes TinyAya-Global for balanced performance and three specialized regional variants named Earth, Fire, and Water. Each specialized model deepens capabilities for specific linguistic clusters, such as Africa or South Asia, while maintaining broader multilingual support. This structure provides researchers with the flexibility to choose between global breadth and regional depth depending on their specific needs.

Technical innovations in tokenization allow Tiny Aya to process non-Latin scripts with significantly fewer tokens than traditional models. This architectural choice reduces memory usage and speeds up inference, making it practical for real-world use on modern smartphones. By focusing on data quality over quantity, the models outperform larger competitors in diverse linguistic tasks and mathematical reasoning.

Running AI locally ensures that sensitive data remains on the device, providing a privacy-first solution for healthcare and education sectors. A university lab in India or a remote clinic in Nigeria can now deploy sophisticated translation tools without relying on external APIs. This democratization of technology represents a shift toward a more inclusive and decentralized AI ecosystem.

Cohere has released the weights for Tiny Aya on platforms like Hugging Face and Kaggle. Detailed technical documentation is also provided to help the global research community replicate and build upon these findings. The initiative encourages a vibrant ecosystem where many voices shape the future of machine learning.