France just entered the frontier AI race with a model trained on 4,000 GPUs in European data centers. The weights go public in three weeks. If it works, it rewrites the map.

On Tuesday, French AI lab Mistral released a public preview of Mistral Large 4, a 1-trillion-parameter multimodal model nicknamed Le Chonk.

It’s the largest model Mistral has ever built, with 49 billion active parameters at inference. It speaks over 160 languages and handles text, images, code, and structured data.

The model is available now via API on Mistral Studio. Open weights are planned for October 27.

That three-week gap is deliberate.

Mistral VP Science Pierre Stock told TechCrunch the company is running safety testing and working with “trusted partners and governments to make sure that the open source weights can be used to defend, but not to perform malicious attacks.”

In other words: Mistral knows what happened when Anthropic released Fable without enough testing. It’s not repeating that mistake.

Macron’s “Third Way” Has a Model Now

French President Emmanuel Macron described Mistral’s approach as “a third way in AI.” The pitch: not American closed-source, not Chinese open-weight, but European open-weight with sovereign infrastructure.

That framing has real substance behind it now.

ML4 was trained entirely on Mistral’s own compute, using 4,000 Nvidia Grace Blackwell GPUs deployed in European data centers over two months.

Stock said that’s “two to three times less than our Chinese competitors, and significantly less than the closed source competitors.”

For enterprises and governments worried about where their AI runs, who controls the weights, and which government might pull the plug at any moment, the sovereign angle isn’t marketing. It’s the product.

The growing dominance of Chinese open-weight models has created a vacuum for Western organizations that want open models but aren’t comfortable relying on Chinese infrastructure. Mistral is positioning Le Chonk to fill exactly that gap.

What It’s Built For

Mistral isn’t trying to be everything to everyone. ML4 is optimized for specific enterprise use cases where multimodal capabilities add real value.

Cybersecurity is the lead pitch. Mistral’s co-founder and chief scientist Guillaume Lample told CNBC: “The cyber defence capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks.”

Open weights mean defenders can audit the model themselves.

Finance is the second focus. ML4 is designed for document analysis, regulatory compliance, and structured data extraction across multilingual financial documents.

Chip design is the third, and it’s not a coincidence.

Two of Mistral’s biggest backers are ASML, which led its Series C, and Samsung, which led its €3 billion Series D last month at a €21 billion valuation (about $24.4 billion). Both companies design chips. Both need AI that understands chip architecture. Both funded Mistral to build it.

The Benchmarks Aren’t Here Yet

This is the caveat that matters most.

Mistral says ML4 will be “best in class among open-weight models, especially outside of China.” But benchmarks are still pending.

Quartz noted that the model still trails in coding relative to other frontier systems.

And Mistral acknowledged that its reinforcement learning phase hasn’t been completed yet.

That’s a lot of “will be” for a model that’s already live on the API. Early access users and cybersecurity professionals are testing now.

The weights drop on October 27. Somewhere between those dates, independent benchmarks will tell us whether Le Chonk lives up to the name or just fills the room.

Mistral’s previous Le Chat chatbot earned respect for speed and privacy.

Its open-weight Mixtral models proved that smaller European labs could punch above their weight. ML4 is the bet that Mistral can do it at frontier scale.

Why the Timing Matters

Le Chonk arrives one day after Nvidia-backed U.S. startup Reflection AI released its own first open-weight model.

The open-weight frontier is getting crowded fast, with Chinese labs like DeepSeek, Z.ai, and Qwen all shipping capable models at aggressive prices.

Mistral has also been hosting Chinese models on its platform, which raised questions about whether the company was pivoting into becoming an inference provider rather than a frontier lab.

Le Chonk is the answer: Mistral still builds models, not just runs other people’s.

The $24.4 billion valuation puts serious expectations on this release.

If ML4 performs, Mistral becomes Europe’s legitimate answer to the frontier labs. If it doesn’t, the “third way” stays a political slogan rather than a technical reality.

Three weeks until the weights drop. That’s when we find out.