Mistral Large 4: A One-Trillion-Parameter Preview, With Looser Cyber Limits for Defenders Only
Mistral Large 4 enters public preview with 1 trillion parameters and a 1M-token context. Cyber defenders test a looser version before the weights go public.
The Bright Recap
Mistral Large 4 has 1.05 trillion parameters, uses 49 billion of them per request and accepts up to one million tokens of input. Mistral opened a public preview on 6 October 2026. A build with lighter safety limits goes to security teams and government agencies for testing, and the downloadable weights are due by the end of October.
To know more about this topic, read our related articles:
Bright Answers
When will Mistral Large 4 be available to download?
Mistral says the open weights will be released by the end of October 2026, and Reuters reports a release date of 27 October. Developers can use the public preview through Mistral Studio in the meantime.
Who can use the less restricted version of Mistral Large 4?
Mistral is giving security specialists, approved partners and government agencies a build with lighter moderation and stronger cyber skills, so they can attack it on purpose and find weaknesses before the weights become public.
Mistral AI has opened a public preview of Mistral Large 4, a one-trillion-parameter open-weight model that the French company plans to make freely downloadable by the end of October. In the meantime, a less restricted build with stronger cyber skills is reserved for security teams, approved partners and government agencies, who will probe it for weaknesses.
Meet Cantica, the Fin-Tech intelligence system for custom reports to strategise your next business move.
What Mistral released
Large 4 is the newest model in Mistral's top-end Large line and arrives about a month after the company's €3 billion funding round. Developers can already try the preview through Mistral Studio, the company's platform for building on its models. Chief executive Arthur Mensch presented the model at a conference in Abu Dhabi, according to Reuters.
Inside the model
The Large 4 model card lists 1.05 trillion parameters, of which 49 billion work on any single request. That design, called a mixture of experts, sends each task to the parts of the network suited to it, so a request uses only a fraction of the model's full size. Large 4 reads up to one million tokens in one request and carries a vision component, so it handles images as well as text.
Mistral built the model on 3,800 Nvidia Grace Blackwell chips housed in European data centres that the company runs itself. Mistral says Large 4 works in more than 160 languages, every official language of the European Union among them.
Defenders get the looser version first
The terms of early access appear in Mistral's launch announcement. Security specialists, approved partners and government agencies receive a build with lighter moderation for red-teaming, the practice of attacking a system on purpose to find its flaws before others do. Mistral says the weights, the files that let anyone run the model on their own servers, will follow by the end of the month, and Reuters reports a release date of 27 October.
On Mistral's figures, Large 4 sits in the top five worldwide on the Artificial Analysis Cyber Index and ranks first there among open-weight models built outside China. The company also reports a score of 82% on a test of reproducing known software vulnerabilities, ahead of every other model it measured.
Finance among the sectors behind the model
Finance is one of the sectors Mistral names among the enterprises that helped develop Large 4, alongside engineering, manufacturing, logistics and pharmaceuticals. Banks and fintech companies testing the preview can use Mistral's European deployment, which the company says it operates entirely under European law without relying on other digital service providers.
The version with loosened cyber limits stays with the defenders Mistral picked to test it until the weights go public.
Editor's note
Every piece published on The Bright Minded goes through careful verification, but mistakes can happen. If you spot an error, have additional information, or want to flag anything, write to rosalia@thebrightminded.com.