Mistral AI has released its first language model, Mistral 7B, as a free download. The French startup says the model can deliver capabilities comparable to other small language models while requiring less computing power, making it possible for more people and organizations to try running a model themselves.
A smaller model designed to be easier to run
Mistral 7B is positioned as a more compute-efficient option alongside models such as Llama 2. The company’s claim is based on standard benchmarks, according to the source article. That comparison matters because the resources needed to run an AI model can shape who can use it and where it can be deployed.
At the other end of the scale, foundation models such as GPT-4 can do more, but are more expensive and difficult to operate. The source describes those larger systems as generally offered through APIs or remote access. A smaller model that can run locally offers a different path for developers and organizations with suitable hardware, or those prepared to pay for cloud computing.
Free to use, with a permissive license
The model is available for download through several methods, including a 13.4-gigabyte torrent. Mistral also opened a GitHub repository and a Discord channel for collaboration and troubleshooting.
Mistral 7B is released under the Apache 2.0 license. As described in the source, the license allows broad use and reproduction, with attribution as the stated condition. That means the model may be used by individuals, businesses, or the Pentagon, provided they can supply the computing resources to run it.
That freedom can lower the barrier to testing and building with the model. But a free download does not make every part of the system public. The model was developed privately with private funding, and the datasets and weights are also private, according to the article.
Open use does not mean every detail is open
The release highlights a distinction between being able to use a model and having access to how it was made. Mistral 7B can be downloaded and used broadly, but the company has not made its data and weights public. The source article cautions against treating the release as fully open source.
That distinction affects what users can inspect or adapt. The permissive license grants room to use the available model, while private training materials and weights limit what outsiders can learn about its construction. Mistral’s CEO, Arthur Mensch, said that the company’s training and dataset assembly processes were proprietary so far.
A free release alongside paid products
The free model also sits alongside Mistral’s commercial plans. The company’s blog post says its paid offering will be distributed as white-box solutions, with both weights and code sources available. Mistral said it was working on hosted solutions and dedicated deployments for enterprises.
Not every future model will necessarily follow the same release approach. Mensch said some future models would be released under Apache 2.0, while larger models would be made available through an API, presumably paid, rather than as a do-it-yourself download.
For now, Mistral 7B gives users a way to experiment with a model that the company describes as capable for its size and less costly to run than comparable options. The release broadens access to using the model, while Mistral’s paid products and private development details remain part of its business offering.