Mistral Large 4 “Le Chonk” Released: 1-Trillion-Parameter Open-Weight AI Targets Cybersecurity

A massive new artificial-intelligence model has entered the cybersecurity race.

French AI company Mistral has released the weights for:

Mistral Large 4

also nicknamed:

Le Chonk.

The model contains approximately:

1 trillion parameters

and is designed for demanding tasks including:

• cybersecurity;
• coding;
• AI agents;
• multimodal reasoning;
• financial analysis;
• professional work.

What makes the October 29, 2026 release particularly significant is not simply the size of the model.

Mistral Large 4 is:

open-weight.

Organizations can download the model and operate it on their own infrastructure instead of depending entirely on a cloud AI provider.

For cybersecurity teams, that can mean:

• keeping sensitive source code in-house;
• running AI inside private environments;
• customizing the model;
• avoiding dependence on an external API;
• continuing to use the model even if a provider changes access policies.

But the same feature creates a serious security debate.

Once the model weights have been downloaded:

Mistral cannot remotely take them back.

Later safety improvements to Mistral’s managed service also do not automatically update copies organizations have already downloaded.

And Mistral itself acknowledges that increasingly powerful cyber models could potentially provide an:

uplift to real-world attacks.

Here is what Mistral Large 4 is, what “open-weight” actually means, how strong its cybersecurity capabilities appear to be, how much it costs, and why Le Chonk is becoming one of the most important AI-security releases of 2026.

What Is Mistral Large 4?

Mistral Large 4 is Mistral AI’s flagship frontier model.

The company describes it as a:

1-trillion-parameter, natively multimodal AI model.

It combines:

• instruction following;
• reasoning;
• coding;
• agentic capabilities;
• image understanding;
• cybersecurity;
• enterprise workflows.

Mistral originally placed Large 4 into public preview on:

October 6, 2026.

The company then continued reinforcement learning and testing with enterprise and cybersecurity partners.

On:

October 29, 2026

Mistral released the model’s weights publicly.

That means organizations can now download the model and run it independently.

Why Is It Called “Le Chonk”?

“Le Chonk” is Mistral’s informal nickname for Mistral Large 4.

The name reflects the model’s enormous size.

With roughly:

1 trillion parameters

it is considerably larger than many commonly deployed open-weight models.

The nickname is playful.

The underlying technology is not.

Mistral is positioning Large 4 as a serious model for governments, enterprises, developers and cybersecurity organizations that want greater control over their AI infrastructure.

Is Mistral Large 4 Open Source?

The most accurate term is:

open-weight.

That is not necessarily identical to traditional open-source software.

Mistral allows users to download the trained model weights.

The model is released under what the company calls the:

Mistral Ecosystem Open-Weight License.

Mistral describes it as a modified MIT license.

According to Mistral, users can generally:

• use the model;
• modify it;
• fine-tune it;
• deploy it;
• commercialize applications built with it.

There is an important exception.

Businesses that operate a:

Model-as-a-Service

offering need a separate commercial license from Mistral.

So describing Mistral Large 4 simply as “free open-source AI” can be misleading.

Open-weight is the safer description.

Why Open Weights Matter for Cybersecurity

Cybersecurity organizations often work with highly sensitive information.

That can include:

• proprietary source code;
• malware samples;
• internal logs;
• incident-response data;
• network architecture;
• authentication systems;
• vulnerability reports;
• confidential customer information.

Sending that information to a third-party hosted AI service can create:

• privacy concerns;
• regulatory concerns;
• data-residency issues;
• operational dependence.

An open-weight model can instead be deployed:

inside the organization’s own environment.

That can give the organization more control over:

• where data goes;
• who can access the model;
• logging;
• network permissions;
• model customization;
• security policies.

This is part of what Mistral calls:

sovereign AI.

What Is Sovereign AI?

Sovereign AI generally refers to the ability of an organization or country to operate artificial intelligence under its own control rather than depending entirely on another company’s infrastructure.

For example, a government may want to avoid sending sensitive cybersecurity information to an external cloud service.

A bank may need strict control over:

where customer data is processed.

A defense organization may want its AI to operate:

without an external internet connection.

An open-weight model can make that possible.

Organizations can potentially run Mistral Large 4:

• on private infrastructure;
• in a private cloud;
• inside regulated environments;
• on premises.

That is one of Mistral’s central arguments for releasing Le Chonk openly.

How Good Is Mistral Large 4 at Cybersecurity?

Mistral says Large 4 performs particularly strongly on defensive cybersecurity tasks.

The company reports that the final model scored:

51%

on the Artificial Analysis Cyber Index.

That benchmark evaluates whether AI systems can:

• find software vulnerabilities;
• reproduce vulnerabilities;
• verify them;
• patch them;
• avoid breaking existing functionality.

A 51% score should not be interpreted as meaning the model finds 51% of all real vulnerabilities.

It is a benchmark score under a specific evaluation framework.

Mistral says Large 4 can assist with legitimate tasks such as:

• vulnerability identification;
• vulnerability reproduction;
• remediation;
• source-code analysis;
• malware investigation;
• log analysis.

Cybersecurity organizations including Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet and others have participated in evaluations or partnerships around the model.

Is Mistral Large 4 the Best AI Cybersecurity Model?

That claim requires caution.

Mistral describes Large 4 as one of the strongest open-weight models for cybersecurity.

However, independent reporting notes that some competing open-weight systems score higher on individual cyber benchmarks.

Benchmarks also change rapidly.

A model that leads one test today may be surpassed weeks later.

Cybersecurity performance also depends heavily on:

• agent frameworks;
• tools;
• context;
• system permissions;
• prompts;
• human expertise;
• vulnerability type.

So the safer conclusion is:

Mistral Large 4 is one of the most capable open-weight cybersecurity models currently available, based on published benchmark results.

It is not possible to establish from one benchmark that it is universally the best cyber model.

Mistral Large 4 Coding Performance

Coding is another major focus.

Mistral reports that Large 4 scores:

48%

on the Artificial Analysis Coding Agent Index.

According to the company, that puts it roughly alongside leading open-weight competitors while using:

47% fewer tokens per task

than GLM-5.3 in that evaluation.

On Terminal-Bench 4.0, Mistral says the final model scored:

41%.

That represents a:

13-percentage-point improvement

over the earlier public-preview version.

This matters for cybersecurity because modern vulnerability research often requires significant coding work.

An AI security agent may need to:

• inspect source code;
• write tests;
• reproduce bugs;
• modify applications;
• analyze crash behavior;
• propose patches.

Coding and cybersecurity capabilities are therefore increasingly connected.

Three Reasoning Modes

Mistral Large 4 supports three reasoning settings.

Low

Recommended for relatively simple tasks.

Medium

Designed for normal everyday workloads.

High

The default for difficult tasks requiring greater precision and deeper reasoning.

This allows developers to balance:

performance

against:

cost and speed.

A simple classification task may not need the same amount of reasoning as a complicated vulnerability investigation.

How Much Does Mistral Large 4 Cost?

Organizations do not have to self-host the model.

Mistral also provides hosted API access.

Current API pricing announced with the final release is:

Input: $0.70 per million tokens

Cached input: $0.07 per million tokens

Output: $2.10 per million tokens

Pricing can change, so developers should confirm current rates before building a production application.

Self-hosting creates a different cost structure because the organization must provide its own:

• GPUs;
• servers;
• electricity;
• networking;
• deployment infrastructure;
• monitoring;
• engineering.

A 1-trillion-parameter model is not something most ordinary laptop users will realistically run locally at full scale.

Where Can You Use Mistral Large 4?

Mistral says Large 4 is available through several routes.

Users can access it through:

• Mistral Studio;
• Mistral’s API;
• OpenRouter;
• Hugging Face;
• NVIDIA-hosted services;
• Microsoft Foundry.

Organizations wanting maximum control can download the weights and deploy them on their own infrastructure.

Mistral mentions serving frameworks including:

vLLM

and:

SGLang.

The company also offers enterprise customization through:

Mistral Forge.

Can You Download Mistral Large 4?

Yes.

The released model weights are available for download.

That is one of the major differences between Le Chonk and many leading proprietary frontier models.

With a closed model such as a cloud-hosted AI system, the developer usually sends a request to the provider’s servers.

With open weights, organizations can potentially:

possess and operate the model themselves.

That creates substantial independence.

It also creates substantial responsibility.

The Security Risk of Downloadable Frontier AI

Once an open-weight model has been downloaded:

the original provider no longer controls every deployment.

The organization running the model controls:

• network access;
• system permissions;
• safety filters;
• fine-tuning;
• agent tools;
• monitoring.

That is useful for legitimate cybersecurity teams.

It can also create opportunities for misuse.

An attacker could potentially modify or weaken safeguards.

That is why powerful open-weight cybersecurity models create a difficult policy question:

Does giving defenders more powerful tools also give attackers more powerful tools?

The answer can be:

yes to both.

Mistral Says Large 4 Refuses Most Malicious Cyber Requests

Mistral says it has built cyber safeguards into the model.

The company reports an average refusal rate of:

96.6%

on malicious cybersecurity requests drawn from benchmarks including:

• JailbreakBench;
• StrongREJECT;
• AgentHarm.

Mistral describes this as the highest refusal rate it measured among the open models it compared.

But benchmark results are not guarantees.

A 96.6% refusal score does not mean:

96.6% of every possible malicious attack will always be blocked.

Real attackers can:

• change prompts;
• use multi-step interactions;
• fine-tune models;
• modify weights;
• build external tools;
• combine multiple models.

Mistral itself acknowledges that as model cyber capabilities improve, there is potential for misuse.

Mistral Says the Model Could Provide “Uplift” to Real Attacks

This is one of the most important parts of the release.

During the public preview, Mistral says it monitored cyber-related activity.

The company reported that the model refused black-box attacks against live targets.

However, Mistral also acknowledged that it sees:

potential for uplift to conduct real attacks.

That does not mean Le Chonk has already been used to hack large numbers of organizations.

It means Mistral believes increasingly capable AI could potentially improve an attacker’s effectiveness.

The company says it is continuing to strengthen its guardrails.

But Downloaded Models Cannot Automatically Receive New Guardrails

This is where the open-weight model creates a difficult trade-off.

Mistral can improve safety on:

• its hosted API;
• future model releases.

But a copy that someone already downloaded is not automatically replaced every time Mistral updates its safeguards.

Mistral Vice President of Science Pierre Stock confirmed to CSO that later safeguard updates apply to managed services and future model releases rather than automatically changing existing downloaded checkpoints.

This means organizations self-hosting Le Chonk become responsible for maintaining their own security controls.

Open Weight vs. Closed AI: The Cybersecurity Trade-Off

Open-weight model	Closed model
Can often be self-hosted	Usually provider hosted
Greater data control	Provider controls infrastructure
Can be customized deeply	Customization may be restricted
Provider cannot easily revoke downloaded model	Provider can restrict access
Safeguards may be modified by operator	Provider maintains safeguards
Local deployment possible	Often requires internet/API
Organization owns more security responsibility	Provider handles more security layers

Neither architecture is automatically safer.

The correct choice depends on:

• use case;
• data sensitivity;
• infrastructure;
• security expertise;
• regulation;
• operational requirements.

Why Cybersecurity Teams Sometimes Dislike AI Refusals

There is another reason Mistral emphasizes open models.

Security researchers sometimes need to perform actions that look dangerous in isolation.

For example, a legitimate defender might ask AI to:

• reproduce a vulnerability;
• analyze malware;
• generate a proof of concept;
• investigate memory corruption;
• inspect suspicious code.

A general-purpose AI safety system could interpret such tasks as malicious and refuse.

Mistral argues that cybersecurity teams need AI systems capable of distinguishing:

authorized defensive research

from:

harmful attacks.

That problem is difficult.

The same technical task can be:

legitimate in a controlled laboratory

and:

illegal when performed against another person’s infrastructure.

Context and authorization matter.

A Specialized Cybersecurity Version Is More Restricted

Interestingly, Mistral is not releasing every cyber capability without restrictions.

The company says selected enterprise design partners are testing a specialized Large 4 variant optimized for advanced cybersecurity work including:

• penetration testing;
• vulnerability assessment;
• red teaming.

Access to that version is:

vetted

and:

contractually restricted.

Testing is limited to systems that participants:

own or are explicitly authorized to test.

This distinction suggests Mistral itself recognizes that some cyber capabilities deserve stronger access controls.

How Does Le Chonk Connect to the AI Cybersecurity Race?

Mistral Large 4 arrives during an extraordinary period for AI security.

Elite Era Trends has recently covered several examples.

AI Built a $250,000 Chrome Exploit

XBOW researchers used AI agents powered by GPT-5.6 Sol to help assemble a Google Chrome exploit chain that earned a:

$250,000 bug bounty.

That demonstrated how frontier AI can accelerate advanced vulnerability research.

Read our article:

AI Built a $250,000 Chrome Exploit in 6 Days.

AI Agents Targeted More Than 1,200 Websites

A separate campaign showed the offensive side of agentic AI.

Autonomous agents were used to investigate more than:

1,200 websites

and gain access to sites associated with more than:

100 companies.

Read:

AI Agents Hacked 100+ Company Websites.

PHP2Shell

AI also played a major role in recent PHP vulnerability research.

Frontier models helped researchers analyze serious vulnerabilities in the PHP runtime used by platforms such as WordPress.

Read:

PHP2Shell Explained.

Together, these stories point toward the same larger trend:

AI is reducing the time and expertise required for difficult cybersecurity tasks.

Why Mistral Large 4 Could Be Important for Small Security Teams

Historically, advanced vulnerability research required highly specialized experts.

AI cannot replace that expertise.

But it can potentially amplify it.

A smaller security team could use models like Large 4 to assist with:

• reviewing code;
• triaging alerts;
• summarizing logs;
• prioritizing vulnerabilities;
• drafting detection rules;
• analyzing unfamiliar software;
• suggesting patches.

That could allow human analysts to spend more time on:

judgment and verification

rather than repetitive analysis.

Could Criminals Download Mistral Large 4?

Because the model weights are publicly available, access cannot be controlled in the same way as a conventional hosted API.

That is one of the central concerns surrounding open-weight frontier models.

However, downloading a capable AI model does not automatically give someone:

instant hacking ability.

Real-world cyber operations still depend on factors including:

• infrastructure;
• vulnerabilities;
• credentials;
• technical knowledge;
• access;
• operational security;
• target configuration.

The risk is that AI may lower some of those barriers.

That is why open-weight cybersecurity models are considered:

dual-use technology.

They can benefit defenders while also potentially being misused.

Is Open-Weight AI More Dangerous?

There is no simple answer.

Open models provide important benefits.

Researchers can:

• inspect behavior;
• test weaknesses;
• build customized safeguards;
• conduct independent evaluations;
• deploy privately.

Closed models give providers more control.

Providers can:

• revoke access;
• monitor misuse;
• update safeguards centrally;
• restrict dangerous functionality.

The security debate therefore involves a trade-off between:

control by the user

and:

control by the model provider.

Mistral strongly favors giving organizations greater control.

Critics argue that irreversible release of highly capable cyber models can also create permanent misuse risks.

Both concerns are legitimate.

Is Mistral Large 4 Better Than GPT-6 Astra?

There is not enough independent evidence to declare a universal winner.

Different models excel at different tasks.

Mistral reports particularly strong performance for Large 4 in:

• open-weight coding;
• cybersecurity;
• enterprise workflows;
• multimodal tasks.

GPT-6 Astra is a proprietary frontier model designed for advanced reasoning, coding, research, computer use and complex professional workflows.

Benchmark scores between them can also depend heavily on:

• task;
• reasoning setting;
• agent framework;
• tool access;
• evaluation methodology.

A more useful distinction is deployment.

Mistral Large 4

Can be downloaded and self-hosted.

GPT-6 Astra

Is accessed through OpenAI-controlled products and services.

For organizations prioritizing private deployment or sovereign infrastructure, that difference can matter more than a small benchmark advantage.

Is Mistral Large 4 Better Than Claude Opus 5.5?

Again, there is no universal answer.

Mistral reports that in one blind coding preference evaluation, Large 4 ranked behind:

Claude Opus 5.5

while outperforming several leading open-weight models.

That suggests Claude remains extremely competitive for coding quality.

But Claude is a proprietary model.

Large 4 offers a capability Claude does not:

organizations can possess and self-host the model weights.

That makes the two products attractive to different types of users.

What Does “Multimodal” Mean?

Mistral Large 4 is:

natively multimodal.

That means it can work with more than text alone.

Mistral says it can reason across:

• images;
• charts;
• documents;
• technical drawings;
• visual data.

This can matter in cybersecurity.

Security teams often deal with information spread across:

• dashboards;
• screenshots;
• network diagrams;
• PDFs;
• technical documentation.

A multimodal model can potentially analyze several of these formats within the same workflow.

Mistral Large 4 Supports More Than 160 Languages

Mistral says the model is natively fluent across:

more than 160 languages.

This reflects Mistral’s broader focus on multilingual and European AI.

For international organizations, multilingual capability can be valuable for:

• global threat intelligence;
• documentation;
• incident reports;
• international teams;
• regulatory work.

Can Mistral Large 4 Run AI Agents?

Yes.

Agentic capabilities are one of the model’s main features.

Instead of simply generating an answer, an agent can potentially:

1. Receive an objective.
2. Decide what steps are needed.
3. Use permitted tools.
4. inspect results.
5. adjust its approach.
6. continue until the task is completed.

That makes Large 4 relevant to the broader shift toward agentic AI.

Read our guide:

What Is Agentic AI? How It Works, Examples & AI Agents vs Generative AI.

Should Companies Give Mistral Large 4 Autonomous Security Access?

Not without strong controls.

An AI security agent can potentially interact with:

• source code;
• credentials;
• servers;
• cloud infrastructure;
• vulnerability scanners;
• ticketing systems.

That makes excessive permissions dangerous.

Organizations should consider controls including:

• least-privilege access;
• network isolation;
• sandboxing;
• human approval;
• detailed logging;
• credential separation;
• restricted tools;
• controlled test environments.

Open-weight AI provides more control.

It also gives organizations more responsibility for deploying that control safely.

Mistral Large 4: Key Facts

Feature	Mistral Large 4
Nickname	Le Chonk
Developer	Mistral AI
Public weights released	October 29, 2026
Parameters	About 1 trillion
Model type	Open-weight multimodal AI
Main strengths	Coding, cybersecurity, reasoning, agents
Cyber Index score reported by Mistral	51%
Coding Agent Index	48%
Terminal-Bench 4.0	41%
Malicious cyber refusal rate reported by Mistral	96.6%
API input price	$0.70 / 1M tokens
API cached input	$0.07 / 1M tokens
API output price	$2.10 / 1M tokens
Self-hosting	Yes
Downloadable weights	Yes
Languages	160+

Benchmark figures above are Mistral-reported results and should not be interpreted as guarantees of real-world performance.

Frequently Asked Questions

What is Mistral Large 4?

Mistral Large 4 is a roughly 1-trillion-parameter multimodal AI model developed by French AI company Mistral.

What is Le Chonk?

Le Chonk is Mistral’s informal nickname for Mistral Large 4.

When was Mistral Large 4 released?

The model entered public preview on October 6, 2026.

Mistral released the improved model weights on:

October 29, 2026.

Is Mistral Large 4 open source?

The more precise description is:

open-weight.

The weights can be downloaded under Mistral’s modified MIT-style open-weight license.

Can I download Mistral Large 4?

Yes.

Mistral makes the model weights available for self-hosted deployment.

Can businesses use it commercially?

Mistral says the model can generally be used, modified, fine-tuned and commercialized.

Businesses providing the model itself as a Model-as-a-Service need a separate commercial license.

How many parameters does Mistral Large 4 have?

Approximately:

1 trillion parameters.

Is Mistral Large 4 designed for cybersecurity?

Cybersecurity is one of its primary advertised strengths.

Mistral has tested it with security organizations for tasks including vulnerability analysis and remediation.

What is its cybersecurity benchmark score?

Mistral reports:

51%

on the Artificial Analysis Cyber Index.

Can Mistral Large 4 help hackers?

The model is intended primarily for legitimate and defensive uses, and Mistral reports strong refusal behavior on malicious requests.

However, Mistral itself acknowledges potential cyber misuse as capabilities improve.

Can Mistral remotely disable downloaded copies?

Once model weights are downloaded and self-hosted, Mistral does not have the same centralized control it has over its hosted API.

Do downloaded copies automatically receive new guardrails?

No.

Later safeguards added to Mistral’s managed service do not automatically update an already-downloaded model checkpoint.

How much does Mistral Large 4 cost?

Mistral’s API launch pricing is:

$0.70 per million input tokens

and:

$2.10 per million output tokens.

Cached input is priced at:

$0.07 per million tokens.

Can it run locally?

The weights can be self-hosted, but the model’s enormous size means practical full-scale deployment requires significant computing infrastructure.

Is Le Chonk better than ChatGPT or Claude?

There is no universal winner.

Mistral Large 4’s biggest differentiator is that its weights can be downloaded and privately deployed.

Bottom Line

Mistral Large 4 represents an important change in the AI cybersecurity race.

The model contains approximately:

1 trillion parameters.

It is designed for:

• cybersecurity;
• coding;
• reasoning;
• AI agents;
• multimodal professional work.

And unlike many frontier competitors:

the weights can be downloaded.

That allows governments, enterprises and security organizations to run the model on infrastructure they control.

Mistral reports a:

51% Artificial Analysis Cyber Index score

and:

96.6% refusal rate on malicious cyber benchmark prompts.

But the company also acknowledges that increasingly capable AI models could provide an uplift to real attacks.

That creates the central question surrounding Le Chonk:

Does open access make cybersecurity stronger—or make powerful cyber capabilities easier to misuse?

The answer may ultimately be:

both.

Open-weight models give defenders more independence, privacy and control.

But once a powerful model is released, access becomes difficult—or impossible—to withdraw.

Mistral Large 4 therefore represents more than another AI model launch.

It is a test of a larger idea:

whether frontier-level cybersecurity AI can remain open while still being deployed responsibly.