Cohere launches a family of open multilingual models

At the ongoing India AI Summit, Cohere, an enterprise AI innovator, has introduced a groundbreaking suite of multilingual models named Tiny Aya. These models are open-weight, allowing developers and researchers to access and modify the underlying code freely. Supporting over 70 languages, Tiny Aya is designed to operate seamlessly on standard devices such as laptops without needing an internet connection. Launched by Cohere Labs, the research division of the company, these models cater to a variety of South Asian languages, including Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. The core model boasts an impressive 3.35 billion parameters, indicative of its advanced capabilities. Additionally, the company has released TinyAya-Global, a specialized version optimized for better adherence to user commands, making it suitable for applications requiring extensive language support. Cohere has further diversified its offerings with regional variants: TinyAya-Earth, focusing on African languages; TinyAya-Fire, emphasizing South Asian languages; and TinyAya-Water, designed for Asia Pacific, West Asia, and Europe. The company emphasized that this tailored approach enhances linguistic accuracy and cultural relevance, resulting in AI systems that resonate more naturally with the communities they serve. The models were trained using a cluster of 64 Nvidia H100 GPUs, ensuring they can operate effectively with limited computational resources. This is particularly beneficial for developers creating applications for linguistically diverse regions such as India, where offline functionality is crucial. The ability to run these models directly on devices paves the way for offline translation applications, expanding their usability without the reliance on constant internet access. The Tiny Aya models are now accessible on HuggingFace, a well-known platform for AI model sharing, as well as the Cohere Platform. Developers can easily download them from HuggingFace, Kaggle, and Ollama for local deployment. To support further research and development, Cohere is also releasing training and evaluation datasets on HuggingFace, along with a technical report detailing its training methods. CEO Aidan Gomez has indicated that the company is considering going public in the near future. As reported by CNBC, Cohere concluded 2025 with impressive financial results, achieving $240 million in annual recurring revenue and experiencing a remarkable 50% growth quarter-over-quarter.