This paper introduces Mistral 7B, a new 7-billion-parameter language model designed for both superior performance and efficiency. The paper highlights how Mistral 7B outperforms larger existing models like Llama 2 (13B) and Llama 1 (34B) in various benchmarks, including reasoning, mathematics, and code generation, while maintaining efficient inference. This is achieved through architectural innovations such as grouped-query attention (GQA) for faster inference and sliding window attention (SWA) for handling longer sequences with reduced computational cost. Furthermore, a fine-tuned version, Mistral 7B – Instruct, demonstrates strong performance in instruction following and human evaluations, showcasing its adaptability and potential for real-world applications, including content moderation.

Ver todos los episodios