Mistral-7B-Instruct

Mistral-7B-Instruct: The 7.3 billion parameter instruction-tuned model; it outperforms Llama 2 13B on all benchmarks, delivering superior performance with half the size.

This is the instruction-tuned version of the 7.3B parameter Mistral model, designed for high-performance chat and instruction-following. It consistently surpasses larger models: specifically, it beats Llama 2 13B across all evaluated metrics and approaches CodeLlama 7B on code tasks. The architecture integrates Grouped-Query Attention (GQA) and Sliding Window Attention (SWA), ensuring faster inference and efficient handling of longer sequences. Released under the permissive Apache 2.0 license, this model is a top-tier, highly efficient open-source solution for deployment.

https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3