Mistral Releases Large 4 Preview, Claiming a Cybersecurity Edge Over Chinese Rivals

Mistral released a preview of Mistral Large 4 on October 6, 2026, putting a name and a trillion-parameter model behind its latest competitive pitch. Earlier that day, CEO Arthur Mensch said its newest system surpassed Chinese models in some areas, including cybersecurity. His comparison came without named rivals or benchmark scores.

Also nicknamed Le Chonk, the model is intended to compete with leading general-purpose systems while targeting specialized business work. Mistral calls it the strongest open-weight model developed outside China by a substantial margin. Open-weight models make their underlying parameters available so organizations can modify and run them themselves, rather than depend entirely on a provider’s closed service.

A cybersecurity claim without a scorecard

Speaking at Ai Everything in Abu Dhabi, Mensch described the newest model as “above the Chinese models on certain aspects, including cyber.” Startup Fortune, citing Reuters, said he supplied no model name, competing system, benchmark or score. His statement was narrower than a claim of overall superiority: cybersecurity was one example of the areas he highlighted.

The launch carries a broader capability claim. Mistral told WIRED that Large 4 is “very, very close” to some proprietary models. That does not make it a leader in every task. CNBC reports that it still trails the leading systems in areas including coding, even though coding is one of its stated specialties.

Preview access is not the final release

Immediate access is described differently by the two outlets. CNBC says the preview is going to developers, cybersecurity leaders and state authorities, ahead of a wider release later in October. WIRED describes the model as freely available for anyone to use and customize, with a final version expected by the end of the month.

CNBC also describes distribution of the model’s underlying weights as forthcoming. That leaves a material distinction between using the preview and obtaining the parameters needed to run and customize the model independently.

If you use a closed model, there is no guarantee it will still be there tomorrow.

Guillaume Lample, Mistral cofounder and chief scientist, speaking to WIRED

Specialized work and control over defenses

Mistral says Large 4 is optimized for cyberdefense, coding, manufacturing, finance and electrical engineering. Lample told WIRED that specialized domains offer opportunities other labs may not emphasize. He also argues that relying on proprietary models to repel cyberthreats exposes companies to the risk of losing access to a critical defense.

According to Mistral, Large 4 was trained for two months on 4,000 Nvidia Grace Blackwell GPUs in its own European data centers. The company also told WIRED it trained the model from scratch.