Mistral releases trillion-parameter ML4 weights for cyber defense

By Sophie Lindqvist

Mistral has released the downloadable weights for Mistral Large 4 (ML4), a trillion-parameter open-source AI model built for cybersecurity. The French company first introduced the model earlier this month under the nickname “Le Chonk.”

Businesses and government agencies can obtain ML4 through Hugging Face or Mistral’s API and deploy it for their own purposes. For Mistral, the pitch is control over where AI runs, not just access through an outside provider. Cybersecurity makes that argument concrete.

What did Mistral actually release?

Mistral calls ML4 both open-source and open-weight. In the company’s description, the weights are “the numerical training that animates the system.” Publishing those values lets other companies and government agencies deploy the model themselves instead of depending on an outside AI company.

Mistral CEO Arthur Mensch said AI should “be a tool for empowerment, not dominance.” Under that principle, he said, Mistral can deliver “state-of-the-art performance” at a “fraction of the compute and capital committed by closed labs.”

As TechCrunch reported, the release comes as Mistral and other open-weight developers compete with closed-model providers OpenAI and Anthropic. Mistral’s argument is that businesses and governments should not need an outside AI company for access to advanced capabilities.

Takeaway: The release gives qualified users a downloadable model while making independence from closed AI services Mistral’s main commercial pitch.

Why does Mistral think open weights matter for cyber defense?

Mistral said in a Thursday statement: “Amid growing cyber threats, ML4 brings state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally.” The company added: “Enterprise and public sector institutions can now download, deploy, and own frontier AI to detect, analyze, and respond to threats across code and IT infrastructure.”

Pierre Stock, Mistral’s Vice President of Science, told TechCrunch that releasing the weights gives organizations more control over how they use AI in defense. His explanation centered on uninterrupted access to the model.

Stock said: “Concretely, it means that, for instance, as a business company or as a government, you’re able to have access to frontier intelligence to defend yourselves against cyberattacks. Access that can’t be cut off by definition because the model is open-weighted,” he said.

In practical terms, Stock’s claim is that defenders can access ML4 without depending on an outside AI company. That control covers both the model’s use and the environment in which it runs.

Takeaway: Open weights let security organizations retain access and deployment control, which Mistral sees as essential for cyber defense.

How does ML4 compare with GLM-5.3?

Mistral says ML4 performs comparably with the Chinese open-weight model GLM-5.3, with one especially concrete efficiency claim: ML4 uses 47 percent fewer tokens per task. The company says that makes ML4 significantly more economical than its competitor.

Mistral described GLM-5.3 as previously considered one of the world’s best open-weight models, particularly for cyber work. That makes GLM-5.3 the direct yardstick for Mistral’s argument that open-weight systems can match competitive performance while requiring less compute.

Mistral also says ML4 outranks every other open model on agentic coding tasks in Artificial Analysis’ Coding Agent Index. The two claims address different buying criteria: cyber performance and token efficiency on one side, coding-agent performance on the other.

In market terms, Mistral is trying to win on operating cost rather than merely publish another model. The 47 percent token reduction is the company’s clearest numerical case for that pitch.

Takeaway: Mistral is claiming competitive performance against GLM-5.3 while making lower token use its clearest economic advantage.

What can security teams do with the model?

Stock told TechCrunch that the weights give organizations more control over how they use AI to defend themselves. He also described practical deployment choices that matter to security teams.

• Access: Mistral says the weights are available on Hugging Face and through its API.
• Deployment: Stock said companies can deploy ML4 on-premises, inside their own infrastructure and ecosystem.
• Threat analysis: Mistral says enterprises and public-sector institutions can use ML4 to detect, analyze, and respond to threats across code and IT infrastructure.
• Incident response: Stock said the model can “identify, coordinate, and respond in real time to major security incidents.”

Taken together, Stock’s described options let an organization choose where the model runs and which defensive tasks it handles. They also remove the need to send every defensive request to a closed provider.

Takeaway: Open weights make control and local deployment central features of ML4, not optional extras.

What did Stock say about rogue AI agents?

Stock acknowledged recent incidents involving rogue AI agents without minimizing the risk. He said the issue deserved attention without treating it as hopeless.

“We don’t want to fearmonger people. We think it’s a problem that is solvable. It’s a hard problem to have good security,” he said.

At the same time, Stock said advances in AI cybersecurity help security professionals find bugs at a much higher rate than before. He wants that capability distributed quickly: “provide [this same] capability to everyone…as soon as possible to ramp up and to strengthen the ecosystem.”

Stock paired that call with a warning that “some malicious actors might try to actually leverage their own technology to attack.” His answer is not to withhold defensive tools, but to expand and strengthen the defensive ecosystem faster.

Takeaway: Stock treats rogue-agent risk as a solvable security problem that wider defensive capability can help contain.

What does the release mean for the open-model market?

In Mistral’s argument, the larger contest reaches beyond cybersecurity. OpenAI and Anthropic are closed-model providers, while Mistral positions ML4 as a downloadable alternative for organizations that want control over advanced AI. That access can determine who sets the operating limits and how defenders respond to incidents.

The model arrives with claims built for adoption: comparable performance with GLM-5.3, 47 percent fewer tokens per task, and the top spot among open models on Artificial Analysis’ agentic coding index. Those claims put pressure on closed-model providers while giving Mistral a measurable pitch to security buyers.

For Mistral, the next test is whether businesses and government agencies download and deploy a trillion-parameter system rather than merely admire its benchmark position. For defenders, Stock’s warning is the enduring constraint: widening access to cyber capabilities must not give malicious actors an easier path to use their own technology offensively.

Takeaway: Mistral’s next challenge is turning its performance and efficiency claims into real deployments while keeping the defensive ecosystem ahead of attackers.