Mistral 7B Instruct is a 7-billion-parameter language model optimized for conversational AI and instruction following. Developed by Mistral AI and released in September 2023, it is a fine-tuned version of the Mistral 7B foundation model. It was designed to demonstrate that smaller models could achieve competitive performance relative to significantly larger counterparts, outperforming several 13B and 34B models on key reasoning and English-language benchmarks at its launch.\n\nThe model is built on a transformer architecture that incorporates efficiency-focused innovations such as Grouped-Query Attention (GQA) and Sliding Window Attention (SWA). GQA reduces inference latency and memory requirements by grouping queries, while SWA allows each layer to attend to a specific window of previous tokens, enabling the handling of sequences with reduced computational overhead.\n\nSubsequent updates to the model, such as version 0.2 and 0.3, expanded the native context window and introduced features like function calling and an expanded vocabulary. The model is released under the Apache 2.0 license, promoting wide accessibility for research and commercial applications.