Cohere Releases Open-Weight Multilingual AI Models
AI company Cohere has launched a family of open-weight models that support over a dozen languages. The release is aimed at democratizing access to enterprise-grade AI. This move continues the trend of major AI labs providing more open access to their foundational technologies.
- The new model family, named "Aya," is an open-weight research release from Cohere's non-profit lab and is available in 8-billion and 35-billion parameter versions. A separate lightweight version called "Tiny Aya" has 3.35 billion parameters and is designed to run on everyday hardware with limited connectivity. - The main Aya models support 23 languages, including German, French, Spanish, Japanese, and Chinese, while the "Tiny Aya" variant covers over 70 languages with a focus on those often underserved by existing models. - In performance benchmarks, Cohere claims the 35B Aya model outperforms larger models like Meta's Llama 3.1 70B, and the 8B version shows advantages over similar-sized models like Google's Gemma 2 9B. - This release is part of a broader initiative to address the English-centric bias prevalent in AI, aiming to improve performance and accessibility for nearly half of the global population. - The model weights and training datasets are available on platforms like Hugging Face, Kaggle, and Ollama, allowing developers to deploy them locally. - Cohere was co-founded by CEO Aidan Gomez, who was one of the eight authors of the original 2017 Google research paper "Attention Is All You Need," which introduced the foundational Transformer architecture used in most modern LLMs. - The Toronto-based company was valued at $5.5 billion as of July 2024 and reportedly closed 2025 with $240 million in annual recurring revenue.
