Aya Expanse

Aya Expanse comes out of Cohere Labs' Aya initiative, an open multilingual research effort with thousands of collaborators, and ships open weights at 8B and 32B parameters covering 23 languages including Arabic, Chinese, Hindi, Japanese, Korean, and Vietnamese. Context windows differ by size: 128K tokens on the 32B model and 8K on the 8B. The models combine several research contributions: data arbitrage for sourcing synthetic multilingual training data, multilingual preference training, safety tuning, and model merging, applied on top of Cohere's Command family architecture. Cohere reports the 32B model winning head-to-head multilingual comparisons against larger open models including Llama 3.1 70B, Mixtral 8x22B, and Gemma 2 27B. Weights are on Hugging Face under CC-BY-NC-4.0 plus Cohere Lab's Acceptable Use Policy, so commercial use is off the table, but research, prototyping, and evaluation are permitted. There is no companion GitHub repository; the models load through standard Transformers tooling and quantized community builds. The wider Aya family now also spans Aya Vision and the compact Tiny Aya models.