By Hamza Faisal

French AI startup Mistral has released its Large 4 model to the public, calling it its most capable openly available model for cybersecurity work. Unlike leading systems such as Google's Gemini, which are mainly reached through online services, Large 4 can be downloaded and run independently by anyone, on their own hardware.

The company says the model is designed to help organisations identify and fix cybersecurity vulnerabilities — finding weaknesses in software before attackers do, and speeding up investigations once a breach has happened. It is the latest move in Mistral's bet that open-weight models, freely downloadable and modifiable, can compete with the closed systems of the biggest US labs.

An upgrade after real-world testing

Mistral said the publicly released version is an upgrade from the one unveiled three weeks ago, refined after testing with governments and corporate users. That testing period helped improve its performance, the company said, before the model was opened to everyone.

The pitch to organisations is control. Because the software can run on a company's own systems, it reduces dependence on a third-party provider — no data has to leave the building, and no outside service can cut off access. That sovereignty argument has become Mistral's signature, and it resonates particularly with European governments and regulated industries.

Open weights, open questions

Mistral's vice president of science, Pierre Stock, framed open weights as a matter of principle. "I'm not comfortable living in a world in which open weights do not exist," he said. Once a model is out, he added, "it's impossible to contain access" — a feature, in his telling, that protects users from being cut off, even though it also means anyone can use it.

That dual-use tension is the backdrop to the release. Open-weight models have surged in popularity because of their low cost and flexibility, and Chinese models in particular dominate the open field. Mistral is positioning Large 4 as a European answer — a capable open model built under European rules and run on European infrastructure.

The timing also matters. The release lands in a week when AI-driven cyberattacks have dominated the news cycle, with researchers reporting that open models can complete complex multi-stage attack simulations. Mistral's argument is the other side of that coin: the same capabilities, in defenders' hands, find and patch vulnerabilities faster than attackers can exploit them.

Orange signs on first

The first major customer is already lined up. Orange's digital security arm announced on Thursday a partnership with Mistral to equip its products with a cybersecurity variant of the company's latest model. Orange Cyberdefense, one of Europe's largest cybersecurity providers, will rely on computing capacity located in Europe to run models dedicated to its security operations.

The rollout will be gradual, initially limited to select services for corporate and public-sector clients. The initial focus: testing clients' defences to find vulnerabilities before attackers do, and investigating breaches faster once they occur. Orange Cyberdefense reported 1.3 billion euros ($1.45 billion) in revenue in 2025; financial terms of the Mistral deal were not disclosed.

For Mistral, the Orange deal is validation of a strategy it has pursued since its founding: build frontier-class models in Europe, keep the weights open, and sell sovereignty as a feature. Whether Large 4 wins over security teams will depend on what it finds in the wild — but from Thursday, anyone can download it and check.