Cohere Releases Aya 23 Open-Weight Multilingual AI Models in 8B and 35B Sizes

Cohere for AI, the company's non-profit research arm, has launched Aya 23, a family of open-weight, instruction-tuned language models available in 8-billion and 35-billion parameter versions. Unlike its predecessor Aya 101, which spanned 101 languages broadly, Aya 23 focuses on delivering deeper, higher-quality performance across 23 languages — including Arabic, Chinese, Hindi, and Spanish — covering roughly half the global population. The models are built on Cohere's Command series and fine-tuned on the Aya Collection dataset, with weights publicly available on Hugging Face under a CC-BY-NC license. The 8B model is designed to run on consumer-grade hardware, lowering the barrier for developers, while the 35B version benchmarks favorably against open models such as Gemma and Mistral on multilingual tasks. The release offers developers building non-English applications a practical open alternative to closed APIs or the costly effort of fine-tuning English-centric models.