Cohere Launches Open Multilingual AI Models
AI research lab Cohere has released Tiny Aya, a family of open-weight models that support over 100 languages. The release is part of a broader trend of major AI labs providing open-source models to lower barriers for developers building globally accessible AI applications. The models were announced alongside the India AI Impact Summit, which is convening major technology firms like OpenAI, Nvidia, and Google.
- The "Tiny Aya" model has a base size of 3.35 billion parameters, making it small enough to run on devices like laptops without constant internet connectivity. It was trained using a relatively modest 64 Nvidia H100 GPUs. - This release is part of a larger "Aya" project by Cohere For AI, a non-profit research lab. The overarching Aya initiative has involved over 3,000 researchers from 119 countries to address the lack of language diversity in AI. - The models are built on the massive "Aya Collection," a dataset of 513 million prompts and completions covering 114 languages. This collection was created by collaborating with fluent speakers to create templates and by translating existing datasets. - While the base model is multilingual, Cohere has also released regionally-tuned versions: TinyAya-Earth for African languages, TinyAya-Fire for South Asian languages, and TinyAya-Water for Asia Pacific, West Asia, and Europe. - The term "open-weight" means that while the model's trained parameters are publicly available for download and modification, the training code and original datasets may not be fully disclosed, distinguishing it from fully "open-source" releases. - The models are available for local deployment through platforms like Hugging Face, Kaggle, and Ollama, allowing developers to integrate them without vendor lock-in. - The Aya family of models includes larger versions that have demonstrated strong performance, with the 13-billion parameter Aya 101 model outperforming comparable models like BLOOM and mT0. More recent versions, like Aya Expanse 32B, have shown advantages over models like Google's Gemma 2 and Meta's Llama 3.1. - This initiative is a strategic move by Cohere, a company valued at $6.8 billion, to compete with major AI labs like OpenAI and Google by focusing on enterprise-ready, data-sovereign AI that can be deployed in private cloud or on-premise environments.
