Oct 7, 2026 · 9 min · 8 segments
Wednesday morning AI and cyber briefing. Mistral opens a public preview of Mistral Large 4, its biggest model, pitched at security teams, with open weights due by the end of October and a…
And yesterday, it opened a public preview of Mistral Large Four, its biggest model yet.
For now, it's only on Mistral's paid online service, but open weights are due by the end of the month, reportedly on the twenty-seventh.
Open weights means the model files themselves go public, so any organization can download a copy, host it on its own servers, and change it.
It's pitched hard at security teams.
On an independent cyber ranking from Artificial Analysis, a firm that tests AI models, it sits in the top five in the world.
On one test, where a model must reproduce a real flaw in freely available software and then fix it, it scored eighty-two percent, the top score on that test, while some leading closed models score close to zero because they refuse the task.
Until the files go public, vetted security firms, partners, and government agencies get the same model with reduced moderation, meaning the filters that make it refuse harmful requests are turned down, plus extra cyber abilities.
Every score is still preliminary.
This one is for Tim because security teams will ask whether to try it now or wait.
For a European bank, a government department, or anyone who must keep models in-house, that's a real option.
I'd try the preview this month on a few security jobs that don't involve sensitive data, like writing the rules that flag suspicious activity.
Then plan the proper test for when the files arrive because the scores will move, and the version you download may not be the one you tried.
But isn't a version with its safety filters turned down exactly the thing that worries regulators?
It's a fair worry, and it's the part of the announcement with the least detail.
The honest case for it is that refusals can block defenders in the middle of a real incident.
If you're offered that looser version, get the rules in writing on who qualifies, what gets logged, and what you're allowed to do before anyone on your team signs on.
So try it now on safe work and test it properly when the files land.
And yesterday, it opened a public preview of Mistral Large Four, its biggest model yet.
For now, it's only on Mistral's paid online service, but open weights are due by the end of the month, reportedly on the twenty-seventh.
Open weights means the model files themselves go public, so any organization can download a copy, host it on its own servers, and change it.
It's pitched hard at security teams.
On an independent cyber ranking from Artificial Analysis, a firm that tests AI models, it sits in the top five in the world.
On one test, where a model must reproduce a real flaw in freely available software and then fix it, it scored eighty-two percent, the top score on that test, while some leading closed models score close to zero because they refuse the task.
Until the files go public, vetted security firms, partners, and government agencies get the same model with reduced moderation, meaning the filters that make it refuse harmful requests are turned down, plus extra cyber abilities.
Every score is still preliminary.
This one is for Tim because security teams will ask whether to try it now or wait.
For a European bank, a government department, or anyone who must keep models in-house, that's a real option.
I'd try the preview this month on a few security jobs that don't involve sensitive data, like writing the rules that flag suspicious activity.
Then plan the proper test for when the files arrive because the scores will move, and the version you download may not be the one you tried.
But isn't a version with its safety filters turned down exactly the thing that worries regulators?
It's a fair worry, and it's the part of the announcement with the least detail.
The honest case for it is that refusals can block defenders in the middle of a real incident.
If you're offered that looser version, get the rules in writing on who qualifies, what gets logged, and what you're allowed to do before anyone on your team signs on.
So try it now on safe work and test it properly when the files land.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.