Aya Expanse 8B

Aya Expanse 8B is the smaller of two models in the Aya Expanse family released by Cohere for AI in December 2024 — a nonprofit research effort built on three years of multilingual research with over 3,000 researchers from 119 countries. It covers 23 languages and was trained with innovations including 'data arbitrage' (matching the best teacher model per language for synthetic data), multilingual preference training, and model merging. On the m-ArenaHard benchmark translated into 23 languages, Aya Expanse 8B was reported to outperform open-weight peers like Gemma 2 9B and Llama 3.1 8B with win rates above 60%. Weights are openly downloadable from Hugging Face.\n\nThe honest picture: Aya Expanse is licensed CC-BY-NC with Cohere's acceptable-use policy — open weights, but non-commercial research use only, which matters for anyone planning to ship products. Its benchmarks are author-run on a benchmark the same lab created. For non-commercial multilingual research and fine-tuning, though, it remains a strong, openly accessible baseline.