Aya Expanse 8B is an open-weight multilingual large language model from Cohere Labs, designed to perform well across 23 different languages. It is built on top of the pre-trained Command family of models, combining that foundation with a year of dedicated multilingual research. The 8-billion parameter version sits alongside a larger 32-billion sibling release, giving developers a lighter option that still inherits the multilingual focus of the broader Aya Expanse project.

The model was shaped through several documented research techniques: data arbitrage for assembling multilingual training data, multilingual preference training to align outputs across languages, safety tuning to reduce harmful responses, and model merging to combine the strengths of separately trained variants. These ingredients target practical multilingual use cases rather than narrow single-language benchmarks, and the open-weight release makes them available for local experimentation. Independent evaluations have used Aya Expanse 8B for translation workloads, comparing inference runtimes across serving frameworks such as Transformers and vLLM, which signals the model's relevance for production-style multilingual pipelines where latency and serving efficiency matter.