Mistral AI has released a new 8x22B MoE (Mixture of Experts) model, known as Mixtral 8x22b, marking a significant advancement in their open-source large language models (LLMs). This model, boasting 176 billion parameters and a 65k context length, is designed to be fine-tuned on custom datasets. It's accessible for testing in ALPHA via public cloud services and is expected to deliver performance between GPT-4 and Claude Sonnet. The model maintains compatibility with Mistral's previous tokenizer, identical to the Mistral 7b, and introduces a larger base model for enhanced capabilities. It requires approximately 260GB of VRAM for operation, with detailed comparisons to other state-of-the-art models anticipated soon. This release coincides with Google Cloud Next day, highlighting the model's potential for broad application and research in AI.