Arthur Mensch is tired of hearing that Europe has lost the AI race. "The narrative that Europe cannot compete is something that is not true," the Mistral CEO told The Next Web as his company put out a preview of Large 4, a trillion-parameter model it has nicknamed "le Chonk". The surprise is where Mistral chose to fight: the security work that American models won't do.
Mistral is courting security teams that want to dig into software flaws in-house without a US provider's filters saying no. One test in the Artificial Analysis Cyber Index asks a model to recreate a real software bug and then patch it. Mistral says Large 4 scored 82% on it, while Claude Opus 5.5 and OpenAI's GPT-6 Astra "score near zero" because they refuse.
That doesn't make it a free-for-all, according to Mistral. The company says Large 4 still turns down clearly malicious hacking requests more often than other open models do. A less restricted version is going only to vetted security firms, partners and government agencies for now. Those security scores also come from Mistral's own testing, and nobody outside the company has reproduced them yet.
Neck and neck on paper, months behind in practice
Mistral pitches Large 4 directly against Astra, and on its own image-understanding chart the two land neck and neck. Independent testers are less generous. Artificial Analysis gives Large 4 a 38 on its Intelligence Index, well short of the 58 scored by Claude Opus 5.5. That's still a huge leap for a company whose previous Large model barely registered on the same scale.
Developer Simon Willison put it plainly: Mistral is now "maybe about 6 months behind the frontier". For a European lab that's been written off more than once, that gap is closer than many expected. For a customer choosing a model today, it's still a gap.
Open, eventually
Then there's the word "open." You can't download Large 4 yet. Mistral says the weights will arrive "by the end of the month", and The Next Web reports a target date later this month. No license terms have been published, so nobody knows yet what companies will actually be allowed to do with it.
Mensch frames the launch as Europe proving a point, and Mistral trained the model from scratch on Nvidia (NVDA) chips in its own European data centers. But the open-model race it's really in is with Chinese labs, and that one is close. The Next Web found Large 4 edges Zhipu's GLM-5.3 by a hair on a coding test.
What to watch: whether the weights and a license really show up by month's end, and whether outside testers can match those security scores. Here's the argument worth having, though. If a security team says it's defending its own systems, should an AI help it rebuild a real attack, or are the US labs right to say no?
Sources
- 1.Introducing Mistral Large 4 · Mistral AI
- 2.Introducing Mistral Large 4: Le chonk · Simon Willison's Weblog
- 3.Mistral Large 4 is Europe's trillion-parameter answer to US models that refuse security work · The Decoder
- 4.Europe's Mistral launches Large 4 to challenge China's lead in open AI models · The Next Web
- 5.Mistral Large 4 Preview: Intelligence, Performance & Price Analysis · Artificial Analysis
Reported by the WattsUpNext desk from the sources linked below. Spot an error? Tell us at corrections@wattsupnext.com.
The WattsUpNext Brief
Get stories like this in your inbox.
One email each weekday, only the topics you choose. Real news, sourced, no fluff. Unsubscribe in one click.
