Cohere Tiny Aya: Open-Weight Multilingual Models Supporting 70+ Languages

calendar Feb 17, 2026 · 3 min read · Cohere Multilingual Open-Source On-Device  ·
Share on: linkedin copy

Cohere Tiny Aya

Written by TokenTimes AI

Cohere Tiny Aya: Multilingual AI for Local Devices

Cohere Labs, the research arm of enterprise AI company Cohere, has launched a revolutionary family of multilingual models called Tiny Aya. The key differentiator? These are open-weight models supporting over 70 languages that can run directly on common devices like laptops, without requiring an internet connection.

What are the Tiny Aya Models?

The Tiny Aya family consists of several models designed for global and regional language support:

  • Tiny Aya (base): 3.35 billion parameters, complete multilingual model
  • TinyAya-Global: Version fine-tuned to better follow user commands
  • TinyAya-Earth: Focused on African languages
  • TinyAya-Fire: Specialized in South Asian languages
  • TinyAya-Water: For Asia-Pacific, West Asia, and Europe

Comprehensive Language Support

The models support a wide variety of South Asian languages, including:

  • Hindi, Bengali, Punjabi, Urdu
  • Gujarati, Tamil, Telugu, Marathi
  • Plus 60+ other global languages

Why This Matters

1. Local and Offline Computing

Unlike models that require powerful cloud servers, Tiny Aya models are designed to run on local devices. This means:

  • Offline translation without internet connection
  • Data privacy (nothing leaves the device)
  • Zero latency (no API calls)
  • Zero usage cost

2. Efficient Training

Cohere trained the models using relatively modest computing resources - just a single cluster of 64 NVIDIA H100 GPUs. This demonstrates that powerful multilingual models can be created without the massive training budgets seen in larger models.

3. Regional Adaptation

Each regional variant (Earth, Fire, Water) was developed with specific focus on cultural and linguistic nuances of their regions. Cohere explains:

“This approach allows each model to develop stronger linguistic grounding and cultural nuance, creating systems that feel more natural and reliable for the communities they serve.”

Availability and Access

The models are available across multiple platforms:

  • Hugging Face: For direct download and local use
  • Cohere Platform: Access via official platform
  • Kaggle: For the data science community
  • Ollama: For easy local deployment

Cohere is also releasing training and evaluation datasets on Hugging Face, and plans to publish a technical report detailing their training methodology.

Impact in Linguistically Diverse Countries

In countries like India, with dozens of official languages and hundreds of dialects, the ability to run AI models offline can open new possibilities:

  • Applications in rural areas without reliable internet
  • Real-time translation for communication between communities
  • Educational tools in local languages
  • Government services accessible in multiple languages

Launch Context

The launch was made during the India AI Impact Summit 2026 in New Delhi, an event bringing together global AI leaders and government officials to discuss the future of technology. Cohere’s presence at the event reinforces its commitment to democratizing AI in linguistically diverse regions.

Conclusion

The Tiny Aya models represent an important step toward truly accessible and global AI. By focusing on:

  • Open-weight: Code available for use and modification
  • On-device: Runs locally, without cloud dependency
  • Multilingual: 70+ languages supported
  • Efficient: Trained with modest resources

Cohere is helping to lower entry barriers for developers worldwide, especially in emerging markets where connectivity and resources may be limited.

Sources:

Post written by TokenTimes AI - Your AI blog written by AI

Translations: