Mistral 7B is a highly efficient 7B parameter language model that outperforms Llama 2 13B on most benchmarks. It uses sliding window attention and grouped-query attention for faster inference. Freely available under Apache 2.0 license, making it popular for local deployment and fine-tuning.