What changed
On October 29, French AI company Mistral released the weights for Mistral Large 4, a 1 trillion-parameter model nicknamed Le Chonk. With the weights, a business or government can download and run the model on its own hardware instead of calling an external API. Mistral first previewed Large 4 on October 6 and spent three more weeks refining it before the public release.
The model uses a mixture-of-experts design: only about 52 billion of its trillion parameters run at a time, which lowers the computing cost. Mistral says you can run Large 4 on as few as four high-end Nvidia GPUs. The company also released a specialized cybersecurity variant, available only to vetted security partners, for penetration testing and red-teaming.
What you can do with it
If your clinic, law office, or agency handles data you cannot send to a third party — patient records, legal documents, client financials — you can now run a trillion-parameter model on servers you own or lease. That keeps the data inside your perimeter. For firms with in-house technical staff or a managed-services partner, this is the first time a model at this scale has been available to self-host.
The cost is not trivial: four high-end GPUs and the electricity to run them will exceed what most small teams pay for API usage. But for regulated industries or government contractors with strict data-residency requirements, the option to avoid sending documents off-site may be worth the hardware expense. Mistral's license allows internal use but requires a separate commercial agreement if you want to resell the model as a service.
What to keep in your own hands
Self-hosting does not mean zero risk. You still need someone on staff or contract who can secure the server, update the software, and monitor for unauthorized access. The model weights are open, but running them safely is not automatic. If your team lacks the technical skill to manage a GPU cluster, this release does not yet change what you can do — stick with a hosted API until you have the capacity to run it securely.
And just because you can run the model does not mean you own the output. Mistral's license limits commercial resale, so if you plan to build a product around Large 4, read the terms before you ship. Finally, Mistral acknowledged that the model could increase attackers' capabilities in cybersecurity scenarios, even with safeguards in place. If you deploy it, assume that someone else will modify it to remove those safeguards, and plan your defenses accordingly.