The paper introduces Aya, a massively multilingual generative language model that can follow instructions in 101 languages, with over 50% of these languages being lower-resourced. Aya outperforms existing models like mT0 and BLOOMZ on most tasks while covering twice as many languages. The authors have developed extensive new evaluation suites to broaden the state-of-the-art for multilingual evaluation across 99 languages, including discriminative and generative tasks, human evaluation, and simulated win rates. They also conduct detailed investigations on data weighting, pruning, and safety concerns. The Aya model is open-sourced under an Apache 2.0 license, aiming to empower researchers and practitioners to advance multilingual models and applications. Key contributions include expanding language coverage, broadening multilingual evaluation, and addressing data provenance and safety concerns.The paper introduces Aya, a massively multilingual generative language model that can follow instructions in 101 languages, with over 50% of these languages being lower-resourced. Aya outperforms existing models like mT0 and BLOOMZ on most tasks while covering twice as many languages. The authors have developed extensive new evaluation suites to broaden the state-of-the-art for multilingual evaluation across 99 languages, including discriminative and generative tasks, human evaluation, and simulated win rates. They also conduct detailed investigations on data weighting, pruning, and safety concerns. The Aya model is open-sourced under an Apache 2.0 license, aiming to empower researchers and practitioners to advance multilingual models and applications. Key contributions include expanding language coverage, broadening multilingual evaluation, and addressing data provenance and safety concerns.