Cohere releases Aya Expanse 8B and 32B open-weight multilingual LLMs with state-of-the-art cross-language performance

AI Impact Summary

Cohere releases Aya Expanse 8B and 32B open-weight multilingual LLMs that achieve state-of-the-art cross-language performance using data arbitrage, multilingual preference training, safety tuning, and model merging. The open weights and published benchmarks against Gemma 2 9B, Llama 3.1 70B, and Ministral 8B enable rapid evaluation and potential migration for multilingual workloads. Teams should plan to replicate or adapt evaluations on the Arena-Hard-Auto dataset, and incorporate governance and safety checks when deploying across languages.