Mistral releases weights of 1-trillion-parameter Large 4 model, aimed at cybersecurity

Mistral, the French AI startup, released the weights of Mistral Large 4 (ML4) this Thursday (29th), its open-source model with one trillion parameters. Internally nicknamed "Le Chonk," the model had been unveiled in early October 2026, but only now can companies and governments download the weights and run the system on their own infrastructure, without going through Mistral's API or that of any other provider.

The difference is concrete for those deciding on architecture: open weights mean the model can be deployed on-prem, in a private cloud, or in an air-gapped environment, without depending on an external service that can change its pricing, go offline, or, in more extreme scenarios, be cut off by another country's decision. The files are available on Hugging Face, and remain accessible via Mistral's own API for those who prefer not to manage the infrastructure.

The angle is cybersecurity, not general use

Mistral is positioning ML4 specifically as a cyber defense tool, not as a generic competitor to general-purpose assistants. In a statement, the company described the launch this way:

Amid growing cyber threats, ML4 brings state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally. 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.

Mistral, official statement

In practice, this means a model designed for tasks like incident triage, code analysis for vulnerabilities, and automated response to attacks, running within the organization's own perimeter instead of sending sensitive logs and code to a third-party service.

How it compares to other open models

Mistral claims ML4 performs comparably to GLM-5.3, a Chinese open-weight model previously considered one of the best in the world, especially for security tasks. The difference, according to the company, lies in operational cost: ML4 reportedly uses 47% fewer tokens per task than GLM-5.3, making it significantly cheaper to run in production.

The company also claims the model outperforms all other open models in agentic coding tasks on the Coding Agent Index, a benchmark maintained by AI evaluation site Artificial Analysis. This is relevant because it is precisely in agentic workflows, such as an agent scanning a repository, identifying a flaw, and proposing a patch, that the cost per task (measured in tokens consumed) adds up quickly in production.

Mistral CEO Arthur Mensch framed the launch as part of a broader contest between open models and closed providers like OpenAI and Anthropic:

AI should be "a tool for empowerment, not dominance.

Arthur Mensch, CEO of Mistral

What changes for infrastructure decision-makers

Pierre Stock, Mistral's VP of Science, detailed to TechCrunch what opening the weights changes in practice for security and infrastructure teams:

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.

Pierre Stock, VP of Science at Mistral

Stock added that the model can be used to identify, coordinate, and respond in real time to large-scale security incidents, and that opening the weights enables on-prem deployment within the organization's own ecosystem. For an SRE or AppSec team that today depends on API calls to a closed provider, this eliminates two recurring frictions: network latency in incident response and the obligation to route source code or sensitive logs outside the company's perimeter.

The counterpoint: AI is also the attacker's weapon

Stock acknowledged that publicly distributing a frontier model has an uncomfortable side: the same power that helps the defense team also becomes available to attackers. Mistral's position is that this symmetry already exists regardless, and that the answer is to accelerate defense, not hold back the technology.

On recent incidents involving AI agents that acted in unintended ways, Stock avoided an alarmist tone:

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.

Pierre Stock, VP of Science at Mistral

What remains open

Mistral has not publicly detailed the minimum hardware requirements for running a one-trillion-parameter model in production, nor the inference cost beyond the comparative token figures against GLM-5.3. For teams evaluating adoption, this means the real test, how much GPU, how much memory, and what latency an on-prem deployment requires, still needs to be done case by case before deciding between running ML4 locally or consuming it via API.

There is also no direct comparison in the company's statements of total cost between maintaining proprietary infrastructure for a model of this size and simply paying for a managed API. For most teams, the decision will likely depend less on the token benchmark and more on how much regulatory or contractual weight exists to keep security data inside their own datacenter.