Update: It's official. Mistral has announced Mistral Large 4, nicknamed "Le Chonk." It has 1 trillion parameters, with 49 billion active, and it's natively multimodal. Mistral calls it the best open-weights model from the US or Europe on aggregated benchmarks, which notably stops short of claiming it beats Chinese open models overall. It says Large 4 is state of the art for cyber defense, manufacturing and finance work, and beats closed frontier models on visual grounding. It's available through Mistral's API now, open weights are due at the end of October, and Mistral is working privately with cybersecurity partners.

Meet Mistral Large 4, aka Le Chonk.

— Mistral AI (@MistralAI) October 6, 2026

Mistral is about to show off a new AI model, and CEO Arthur Mensch is already making a bold claim about it. Speaking at the Ai Everything conference in Abu Dhabi, he said the model "is actually above the Chinese models on certain aspects, including cyber," Reuters reports.

"So the narrative that Europe cannot compete is something that is not true," Mensch added. He didn't say which Chinese models he meant, which tests he used, or even what the new model is called.

That last part matters. At the time of writing, the model isn't on Mistral's news page or its Hugging Face page, so there's nothing to check the claim against yet.

The best clue to the name comes from Mensch himself. At the Bits & Pretzels founder conference in Munich, he told Handelsblatt that Large 4, Mistral's most powerful model yet, was days away after missing its original summer target. "We don't have a delay. It's a long summer," he joked. Mistral hasn't confirmed that the Abu Dhabi model is Large 4.

If it is, there's a solid base to build on. Mistral Large 3 is an open-weight Mixture-of-Experts model with 675 billion total parameters, 41 billion of them active for each token, released under the Apache 2.0 license. At launch, Mistral said it matched the best open models. It didn't claim to beat them.

Notice who Mensch is comparing against. He isn't claiming to beat OpenAI or Anthropic. Chinese labs behind DeepSeek, Qwen, Kimi and GLM have produced many of the strongest open-weight models, so this is a claim about leading parts of the open-model race, not the whole frontier.

Cyber is also a tough category to win. As we reported, a US government assessment found Z.ai's GLM-5.3 was the strongest open-weight model it had tested on cyber tasks, and Anthropic says it can build working exploits at a level approaching its restricted Claude Mythos Preview.

So "beats Chinese models on cyber" could mean a lot or very little. Topping GLM-5.3 on a serious exploit test would be big news. Edging out an older Chinese model on one narrow quiz would not. Until Mistral names its benchmarks, there's no way to tell.

Why cyber? Because Mistral's customers care about it. When Mistral announced its partnership with Saudi Arabia's HUMAIN, the two companies named cybersecurity and voice as the first focus areas for the models they'll develop together.

Don't confuse this with Shieldstral, which Mistral released this summer. That's a small 3B safety classifier that checks text and images against plain-language content policies. It doesn't hunt for software vulnerabilities.

The Abu Dhabi pitch was as much about geopolitics as tech. Mensch said Mistral's independence from both the US and China is helping it grow quickly in the Gulf and Asia-Pacific, where there's "an enormous need for alternative technology suppliers."

That's also the story behind Mistral's recent €3 billion funding round, led by Samsung at a valuation above €21 billion. Mistral says it now works with more than 125 large companies, including HSBC, Airbus and ASML.

There's one more thing to watch. If the new model is open-weight and as strong at cyber as Mensch suggests, Mistral will walk straight into the debate GLM-5.3 started over freely downloadable hacking skills. Mensch's position is clear: "Only an open ecosystem can guarantee the safety of AI," he posted recently.

Once the model is out, the details will tell the real story: its name, its license, the exact benchmarks, and which Chinese models it was measured against.