aya-expanse-8b

Aya Expanse 8B is an open-weight multilingual large language model developed by Cohere Labs (formerly Cohere For AI). Built on the Command model family and optimized for 23 languages—including Arabic, Chinese, Czech, Dutch, English, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Turkish, Ukrainian, and Vietnamese—it aims to close the performance gap between multilingual and monolingual models in the 8-billion parameter class.

The model uses an optimized autoregressive Transformer architecture with SwiGLU activations, RoPE positional embeddings, and grouped-query attention. Post-training combines supervised fine-tuning with multilingual data arbitrage (strategic sampling from a pool of teacher models), offline and online preference optimization (DPO), safety tuning, and model merging across language-family clusters. It supports an 8K token context window and is released under the CC-BY-NC license.

On academic benchmarks, Aya Expanse 8B scores 57.2 on Global-MMLU and 93.2 on MGSM (multilingual grade-school math). On m-ArenaHard—a multilingual translation of Arena-Hard-Auto—it achieves win rates of 70.6% vs Llama 3.1 8B, 60.4% vs Gemma 2 9B, 55.7% vs Ministral 8B, and 63.1% vs Qwen 2.5 7B, establishing state-of-the-art multilingual performance among comparable open-weight models.