In the ever-evolving world of artificial intelligence, open models are making significant strides. Mistral, a renowned French AI startup, recently unveiled its inaugural model, boasting superior performance for its size. The best part? It's entirely free for everyone to use.
The Mistral 7B model is now accessible for download through various channels, including a 13.4-gigabyte torrent. To foster collaboration and address potential issues, Mistral has also initiated a GitHub repository and a Discord channel.
A standout feature of this release is its licensing under the Apache 2.0 license. This permissive license ensures that anyone, from individual enthusiasts to large corporations, can utilize the model, provided they give appropriate credit.
Mistral 7B stands out in the realm of "small" large language models. While it offers capabilities akin to models like Llama 2, it does so at a significantly reduced computational cost. Although powerhouse models like GPT-4 offer more extensive functionalities, their high operational costs limit their accessibility, often restricting them to API or remote access.
In a recent blog post, Mistral's team expressed their ambition to champion the open generative AI community. They aim to elevate open models to state-of-the-art performance levels. The Mistral 7B model is a testament to their dedication, resulting from three months of rigorous work. This effort involved assembling a top-tier AI team, creating a high-performance MLops stack, and designing an advanced data processing pipeline.
It's worth noting that the founders of Mistral have previously contributed to similar models at tech giants like Meta and Google DeepMind. Their expertise undoubtedly played a pivotal role in this achievement.
However, it's essential to differentiate between "free to use" and "open source." While Mistral 7B is available for everyone, its development involved private funding, and its datasets and weights remain proprietary.
Mistral's business strategy is intriguing. While they offer the model for free, those seeking a deeper dive might be inclined towards their paid offerings. Their commercial products will provide both weights and code sources. Additionally, they are actively developing hosted solutions tailored for enterprise needs.
In a conversation with Mistral's CEO, Arthur Mensch, he hinted that not all upcoming models would be under the Apache 2.0 license, but some would. Larger models might be accessible via a paid API instead of a do-it-yourself method. However, he remained tight-lipped about the training and dataset assembly processes, citing them as proprietary for the time being.