Cohere unveils Tiny Aya, a family of open-source multilingual AI models designed for offline use and edge devices, announced at the India AI Summit.

The Tiny Aya family spans more than 70 languages, with region-specific variants like TinyAya-Earth for African languages, TinyAya-Fire for South Asian languages, TinyAya-Water for Asia Pacific, West Asia, and Europe, and a globally tuned TinyAya-Global that better follows user commands.

The base Tiny Aya model contains 3.35 billion parameters.

TechCrunch notes the timing at the Boston event on June 23, 2026, and references Cohere’s reported 2025 revenue as context for IPO plans.

Future plans include expanding language coverage, optimizing performance on more devices, improving cultural adaptation, and partnering with academic and local developers to align with community needs.

The announcement appears within a sponsor-post context and is part of a broader tech-news roundup on AI tooling.

Availability spans HuggingFace, the Cohere Platform, Kaggle, and Ollama for local deployment, with training and evaluation datasets on HuggingFace and a forthcoming technical report on training methodology.

Training uses a modest setup of 64 Nvidia H100 GPUs, enabling deployment in resource-limited areas, with datasets and the technical report to follow.

Deployment and transparency plans include multiple platforms (HuggingFace, Kaggle, Ollama) for local use and a forthcoming technical report detailing training methods.

The release emphasizes offline capability and broad regional language coverage as part of Cohere Labs’ research efforts.

The models prioritize strong linguistic grounding and cultural nuance while maintaining broad multilingual coverage for flexible customization.

The initiative aims to impact education, healthcare, government services, business localization, and research by providing offline multilingual AI for rural and connectivity-challenged contexts, with a potential market tied to the 1.8 billion speakers of supported languages.