Mistral 7B is a 7-billion-parameter language model designed to deliver high-level performance while maintaining efficient and cost-effective inference. Released under the open-source Apache 2.0 license, the model achieves state-of-the-art results for its size, outperforming the larger Llama 2 13B model across all tested benchmarks and surpassing Llama 1 34B in reasoning, mathematics, and code generation.

The model's efficiency is primarily driven by specific architectural choices:

• Grouped-Query Attention (GQA): This mechanism significantly accelerates inference speed, reduces memory requirements during decoding, and allows for higher batch sizes and throughput.
• Sliding Window Attention (SWA) and Rolling Buffer Cache: SWA allows the model to handle sequences of arbitrary length at a reduced computational cost by having each token attend to a fixed window of previous tokens. Paired with a rolling buffer cache, this approach restricts cache size, reducing memory usage by 8x on 32k token sequences without degrading quality.

Alongside the base model, the authors released Mistral 7B – Instruct, a fine-tuned chat model that surpasses the Llama 2 13B Chat model on both human and automated benchmarks. Furthermore, the model can effectively enforce safety guardrails through system prompts and can perform fine-grained content moderation via self-reflection.