Mixtral is a family of open-weight large language models from Mistral AI built on a sparse mixture-of-experts (MoE) architecture, which activates only a subset of the model's parameters for each input to improve efficiency without sacrificing capability.