Aya (model)
By Cohere For AI
Aya is an open-weight multilingual large language model family developed by Cohere For AI, designed to provide strong instruction-following performance across a much broader set of languages than most contemporary open LLMs, with a…
Definition
Aya is an open-weight multilingual large language model family developed by Cohere For AI, designed to provide strong instruction-following performance across a much broader set of languages than most contemporary open LLMs, with a particular focus on underserved and lower-resource languages.
Overview
Aya was released by Cohere For AI (Cohere's non-profit research lab) in early 2024 as part of a broader open-science initiative to close the multilingual gap in large language model research, which has historically been dominated by English-centric models with only secondary support for other languages. The initial Aya release consisted of two components: Aya Dataset and Aya Collection, massive multilingual instruction-tuning datasets covering 101 languages built through a global crowdsourced annotation effort involving thousands of contributors from around the world, and Aya 101, a 13-billion-parameter instruction-tuned model fine-tuned from the mT5 architecture on that data, which the Cohere For AI team reported outperformed existing open multilingual models on multiple benchmarks across many of those languages. Cohere For AI followed this with the Aya Expanse series in late 2024 (8B and 32B parameter models) built on a more modern architecture, and Aya Vision in 2025, extending the family to multimodal (vision-language) capabilities across dozens of languages. These later releases incorporated techniques like data arbitrage (using multiple teacher models to generate diverse high-quality multilingual training data), model merging, and preference training to substantially improve quality over the original Aya 101, while continuing to prioritize breadth of language coverage relative to comparably sized Western-centric open models. The Aya project is notable both as a model family and as an open-research initiative: the datasets, model weights, and much of the tooling are released openly, and the project has functioned as a rallying point for global NLP researchers, particularly those working on lower-resource languages that receive little attention from major commercial labs. Aya models are positioned less as frontier-capability competitors to GPT-4-class models and more as accessible, high-quality multilingual open-weight alternatives, useful anywhere broad language coverage matters more than raw peak performance in English.
Key Concepts
- Covers 101+ languages, with particular emphasis on underserved and lower-resource languages
- Built on openly released, crowdsourced multilingual instruction datasets (Aya Dataset/Collection)
- Aya Expanse series (8B/32B) improved quality via data arbitrage and preference training
- Aya Vision extends the family to multimodal vision-language tasks
- Fully open-weight releases via Hugging Face under Cohere's open licenses
- Developed by Cohere For AI, a non-profit research lab, as part of an open-science initiative
- Benchmarked against other open multilingual models across broad language sets