Cohere Releases Open Multilingual AI Models
Cohere has launched "Tiny Aya," a family of open-weight models that support over 100 languages. The models are available for both commercial and research use. This release addresses the growing need for AI agents and applications that can operate effectively across different languages and regions.
- The "Tiny Aya" base model has 3.35 billion parameters, making it small enough to run offline on consumer hardware like laptops. It was trained efficiently on a relatively modest cluster of 64 NVIDIA H100 GPUs. - This model family is part of the broader "Aya" project, a massive open science collaboration involving over 3,000 researchers from 119 countries aimed at building AI that covers a much wider range of the world's languages, especially under-represented ones. - To provide better linguistic and cultural nuance, Cohere released specialized regional versions, including TinyAya-Fire (South Asian languages), TinyAya-Earth (African languages), and TinyAya-Water (Asia Pacific, West Asia, and Europe). - The models are available for local deployment via platforms popular with developers and indie hackers, including Hugging Face, Kaggle, and Ollama. - While described as open, the models are technically "open-weight," not fully open-source. The research release, for instance, uses a CC-BY-NC (non-commercial) license, a key distinction for anyone planning to use it in a commercial product. - The models are intended for tasks like multilingual text generation, conversational AI, and translation, demonstrating strong performance in these areas, especially for low-resource languages. However, they are noted to be weaker on complex, chain-of-thought reasoning tasks. - This release is part of a larger industry trend toward smaller, more efficient models optimized for on-device or "edge" deployment, contrasting with the large-scale, cloud-dependent models from providers like OpenAI.
