Aya Expanse 32B is designed as a research-grade multilingual large language model that brings together a high-performing pre-trained Command-family backbone with a year of dedicated multilingual work from Cohere Labs, including data arbitrage, multilingual preference training, safety tuning, and model merging. This pairing lets the model serve 23 languages within a single weights release, rather than relying on separate per-language adapters or narrow translation systems. The intent is to give researchers and builders a capable base for cross-language generation, understanding, and instruction following that holds up across diverse linguistic settings.

Released alongside a smaller 8B sibling, the 32B variant is the stronger option for tasks where multilingual depth, nuanced instruction following, and long-form generation matter most, while still being lightweight enough to self-host compared with much larger frontier systems. Its open-weights availability on Hugging Face and Kaggle makes it a practical fit for academic study, multilingual product prototypes, and community evaluations, particularly for teams that want to fine-tune or audit a capable multilingual model without depending on a closed API. Cohere positions Aya Expanse as outperforming other leading open-weight models on multilingual benchmarks, reflecting the value of combining strong pre-training with targeted multilingual alignment.