Generative models · multilingual research
Language models for the whole world — not just English
Aya is a family of open multilingual research models covering 70+ languages. Built with the global research community, Aya brings high-quality understanding and generation to languages that have long been left behind.
Eval coverage by language
Used by labs and builders working in every language
One model family, every language you work in
Aya was trained to read, reason, and respond across 70+ languages — so quality does not fall off a cliff the moment you leave English.
5/12, 9:15 AM
Aya, translate this support reply into Tagalog and keep it warm.
Harriet 9:16 AM
@Aya how should we adapt this onboarding email for Brazilian Portuguese readers?
Aya APP 9:16 AM
A good multilingual adaptation usually keeps:
- A neutral, fluent rendering rather than a word-for-word transliteration
- Idioms localized so the meaning lands for a native reader
- Tone preserved across the source and target language
Languages covered
0+
Reach beyond the usual few
Aya is trained and evaluated across more than seventy languages, including many that most models barely touch.
Close the language gap, not just the benchmark gap. Aya is built so the next billion speakers get models that actually work for them.
Want the full picture on Aya's multilingual results?
Download the multilingual evaluation reportOpen by design, for the whole research community
Aya is research you can pick up and run with — weights, data, and recipes in the open.
Open weights & datasets
Aya models, the multilingual instruction data, and the evaluation suites are released openly so anyone can study, reproduce, and build on the work.
Fine-tuning recipes
Practical recipes and reference code make it straightforward to adapt Aya to a new language, domain, or task with modest compute.
Responsible-use guidance
Each release ships with model cards, known limitations, and safety notes so teams can deploy multilingual models with eyes open.
“Aya is the first model that handles our regional languages well enough to ship. We went from translating around English to building directly for the communities we serve.”
MERIDIAN
Renata Voss
Lead, Language Technology
Native-speaker review
Outputs are checked by speakers of each language so quality reflects real usage, not just automatic metrics.
Transparent evaluation
Open multilingual benchmarks make it clear where Aya is strong and where a language still needs more work.
Safety across languages
Aya is red-teamed and filtered in many languages, so safeguards are not limited to English alone.
Bias & toxicity testing
Each release is probed for harmful or biased generations across the full range of supported languages.
Documented model cards
Every model ships with a card describing data, intended use, and known limitations for each language family.
Community governance
Researchers worldwide help steer what Aya covers next, keeping the work accountable to the people it serves.
BREVA
How a public health team reaches readers in their first language
A regional health service used Aya to draft guidance in a dozen local languages at once, reaching communities that English-only material had always missed.
Resources for builders and researchers
Everything you need to evaluate, run, and extend the Aya models.