Mistral Says Its New AI Model ‘Le Chonk’ Is the Best Open-Weight Offering Outside of China

Mistral's new one-trillion-parameter model, Mistral Large 4, is in preview now, with a final version due by the end of the month. The company calls it by far the most capable open-weight model built outside China and says it comes very close to some proprietary models. It's optimized for coding and cyberdefense, plus niches like manufacturing, finance and electrical engineering, and Mistral says it trained the model from scratch rather than through distillation.

The timing matters as much as the model. WIRED reports that in June the Trump administration placed temporary restrictions on distributing models from OpenAI and Anthropic over cyberattack concerns, and that the White House has reportedly asked US labs to withhold unreleased models even from the UK's AI Safety Institute. For many businesses and governments, frontier AI access now looks like something that can be revoked.

The line we keep coming back to is from Mistral cofounder and chief scientist Guillaume Lample: what matters is to own the model, because a closed model comes with no guarantee it'll still be there tomorrow. That's a sales pitch. It's also a real argument, and we think it's mostly right.

If a company builds its cyber defenses on top of a model it rents, it inherits every risk that vendor faces — pricing changes, outages, and now export or distribution restrictions. Open weights you can run on your own hardware remove that single point of failure. That's a genuine benefit of useful AI being open, and it's why the sovereignty case is landing in Europe right now.

We'd separate the strategy from the scoreboard, though. Best open-weight model outside China, and nearly level with some proprietary models, are Mistral's claims. WIRED notes that Mistral has generally trailed OpenAI and Anthropic on performance, revenue and release pace, with less capital and compute. A policy researcher quoted in the piece put it bluntly: geopolitics has handed Mistral an opening its performance alone hadn't earned. The claims need independent testing before anyone treats them as settled.

The downside is the flip side of ownership. Once weights are out, nobody can recall them — not Mistral, not regulators. A model tuned for cyberdefense is, by construction, a model that understands offense. If the world splits into blocs that each insist on owning their own frontier models, we get resilience and choice, and we also get many more powerful systems outside anyone's ability to switch off.

I'm watching whether US companies take Lample's pitch seriously. If American enterprises start hedging with European open weights because of their own government's restrictions, that's a meaningful shift in who the AI market trusts.

Mistral earns money by charging pay-as-you-go fees for running models in its cloud and by sending engineers to help customers tune models. In September it raised a $3.3 billion round at a $24 billion valuation, which WIRED describes as the largest raise ever by a European tech company.