Mistral's open-source mixture-of-experts model offering GPT-3.5 performance with efficient inference through sparse activation.