Cohere's Tiny Aya: Open-Weight AI Models That Speak 70+ Languages and Run Offline

Diverse people connected through multilingual AI on smartphones - Cohere Tiny Aya

The AI Revolution Doesn't Start Until It Speaks Your Language

While the biggest names in AI race to top English-language benchmarks, Cohere has taken a radically different approach. At the India AI Summit in New Delhi, the enterprise AI company launched Tiny Aya — a family of open-weight multilingual models that support over 70 languages and can run entirely offline on everyday devices like laptops.

The base model contains just 3.35 billion parameters, making it lightweight enough for local deployment while still delivering meaningful capabilities across dozens of languages — from Hindi, Bengali, Tamil, and Marathi to Swahili, Arabic, and Urdu.

Regional Variants for Global Reach

Cohere didn't just build one model and call it a day. The Tiny Aya family includes purpose-built regional variants:

  • TinyAya-Global: Fine-tuned for broad language support and following user commands
  • TinyAya-Fire: Optimized for South Asian languages
  • TinyAya-Earth: Built for African languages
  • TinyAya-Water: Covers Asia Pacific, West Asia, and Europe

Each variant develops stronger linguistic grounding and cultural nuance while retaining broad multilingual coverage, making them flexible starting points for further adaptation and research.

Trained on Modest Compute

Perhaps the most striking detail: the entire Tiny Aya family was trained on a single cluster of 64 NVIDIA H100 GPUs. In an industry where training runs routinely consume thousands of GPUs and cost hundreds of millions of dollars, Cohere's approach demonstrates that impactful AI doesn't require astronomical budgets.

The models are designed for on-device usage, requiring less computing power than most comparable models. This means developers can power offline translation, local language processing, and more — without any internet connection.

Why This Matters

India alone has 1.5 billion people speaking dozens of languages. Most mainstream AI models barely function in Hindi, let alone Tamil, Punjabi, or Bengali. Across Africa and Southeast Asia, the situation is even more stark. Tiny Aya directly addresses this gap by bringing AI capabilities to communities that have been largely ignored by the AI revolution.

The offline capability is particularly crucial in regions with unreliable internet connectivity, opening up applications in education, healthcare, agriculture, and local commerce that were previously impossible.

Available Now, Everywhere

The models are already available on HuggingFace, Kaggle, and Ollama for local deployment, as well as the Cohere Platform. The company is also releasing training and evaluation datasets, along with a technical report detailing its methodology.

The Bottom Line

With $240 million in annual recurring revenue and 50% quarter-over-quarter growth, Cohere is proving that you don't need trillion-dollar compute budgets to build AI that matters. Tiny Aya is a statement: the most important AI breakthroughs aren't always the flashiest — sometimes they're the ones that bring technology to the people who need it most.