French artificial intelligence company Mistral AI has released the model weights for Mistral Large 4, its trillion-parameter AI model designed for advanced coding, reasoning and cybersecurity tasks. Nicknamed “Le Chonk,” the model is now available for businesses, developers and government agencies to download and deploy on their own infrastructure.
Announced on October 29, 2026, the release follows an initial public preview earlier this month. Mistral says additional reinforcement learning and feedback from enterprise and cybersecurity partners have improved the model’s coding performance, defensive security capabilities and reasoning efficiency. The move also strengthens Mistral’s position in the competition between open-weight AI developers and companies offering access primarily through hosted models.
Key Takeaways
- Mistral released the weights for its trillion-parameter Mistral Large 4 model on October 29, 2026.
- The model is nicknamed “Le Chonk” and is designed for coding, reasoning and cybersecurity.
- It has 1 trillion total parameters and 52 billion active parameters, according to Mistral’s initial technical announcement.
- Businesses and governments can download and deploy the model through Hugging Face.
- Mistral reports a 51% score on the Artificial Analysis Cyber Index.
- The company says the model refuses 96.6% of tested malicious cybersecurity requests.
- Organizations can deploy it on their own infrastructure, giving them greater control over data and security operations.
Mistral Releases Its Trillion-Parameter Large 4 Model
Mistral Large 4 is the company’s latest flagship AI model and one of its largest releases.
The company first introduced the model in public preview on October 6, 2026, promising to release its weights before the end of the month.
It fulfilled that commitment on October 29, publishing an improved version following several weeks of additional training and feedback from developers, enterprise customers and cybersecurity partners.
The model is designed to handle demanding tasks, including software development, complex reasoning, multimodal understanding and security analysis.
Mistral says the latest version delivers improvements in coding and defensive cybersecurity while using fewer tokens for certain tasks.
The release is particularly important because organizations can now access the model’s trained weights instead of depending exclusively on a third-party hosted service.
What Makes Mistral Large 4 Different?
Mistral Large 4 uses a Mixture-of-Experts (MoE) architecture.
Instead of activating every parameter for every request, an MoE model routes individual tasks through selected groups of specialized components, known as experts.
This approach allows a model to contain a very large number of parameters without activating all of them for every token it processes.
According to Mistral’s initial announcement, Large 4 contains approximately 1 trillion total parameters, with around 52 billion active parameters.
The model is also natively multimodal, meaning it can work with different types of information rather than relying exclusively on text.
| Technical specification | Mistral Large 4 |
|---|---|
| Model name | Mistral Large 4 |
| Nickname | Le Chonk |
| Total parameters | 1 trillion |
| Active parameters | Approximately 52 billion |
| Architecture | Mixture of Experts |
| Input capabilities | Natively multimodal |
| Primary applications | Coding, reasoning and cybersecurity |
| Availability | Downloadable model weights and API |
The architecture is intended to balance advanced capabilities with computational efficiency.
Mistral Targets Cybersecurity With Its New AI Model
Cybersecurity is one of the most important areas Mistral is targeting with Large 4.
Security teams frequently need to inspect large codebases, identify vulnerabilities, reproduce software flaws and develop patches before attackers can exploit them.
These tasks can be time-consuming and require specialized technical expertise.
Mistral says Large 4 can assist security professionals with several defensive activities, including:
- Finding security vulnerabilities in software.
- Reproducing and validating identified vulnerabilities.
- Developing patches to fix software weaknesses.
- Analysing suspicious code and potential malware.
- Prioritizing vulnerabilities based on risk.
- Supporting incident response and security investigations.
The company has worked with cybersecurity organizations to evaluate the model in real-world environments.
Its objective is to make advanced AI security capabilities available to businesses and government institutions that need to protect sensitive systems and infrastructure.
Mistral Reports a 51% Score on the Cybersecurity Benchmark
Mistral says Large 4 achieved a 51% score on the Artificial Analysis Cyber Index.
The benchmark evaluates how effectively AI models can discover vulnerabilities in software, reproduce and validate them, and develop patches without breaking existing functionality.
The company also reported that Large 4 achieved an 82% result on a test involving the reproduction and patching of a real software vulnerability.
These results suggest that the model could be useful for security professionals working on complex software environments.
However, the benchmark figures are reported by Mistral and should not be interpreted as proof that the model can independently secure an entire organization.
| Cybersecurity evaluation | Mistral’s reported result |
|---|---|
| Artificial Analysis Cyber Index | 51% |
| Real vulnerability reproduction and patching test | 82% |
| Cybench security challenges | 93% |
| Refusal rate on tested malicious cyber requests | 96.6% |
The figures come from different evaluations and measure different capabilities. They should not be directly compared as if they were results from a single test.
Mistral Says the Model Can Refuse Malicious Cybersecurity Requests
Powerful cybersecurity AI creates an important safety challenge.
The same capabilities that help security professionals identify software vulnerabilities could also be misused by attackers to discover weaknesses and develop exploits.
Mistral says it has built safeguards into Large 4 to distinguish legitimate defensive security work from malicious requests.
According to the company, the model achieved an average refusal rate of 96.6% on malicious cybersecurity prompts from evaluations including JailbreakBench, StrongREJECT and AgentHarm.
This is intended to allow security teams to perform legitimate vulnerability research without unnecessarily blocking their work, while restricting assistance for harmful activities.
However, safeguards are not perfect. The availability of downloadable model weights also means organizations must take responsibility for deployment security, access controls and monitoring.
Businesses Can Run Mistral Large 4 on Their Own Infrastructure
One of the biggest advantages of the release is the ability to deploy the model independently.
Traditional hosted AI services require organizations to send requests to infrastructure operated by an external provider.
For some businesses and government agencies, this creates concerns about data residency, privacy, regulatory compliance and operational dependency.
Mistral’s downloadable weights offer another option.
Organizations can deploy the model within their own infrastructure, including private environments and on-premises systems, subject to the necessary hardware and licensing requirements.
This gives them greater control over how the model is configured, monitored and integrated into existing security systems.
Mistral also argues that open weights reduce the risk of losing access to an AI model because an external provider changes its service or restricts access.
The company has made the weights available through Hugging Face, while organizations can also access the model through Mistral’s API.
Why Open-Weight AI Models Matter for Cybersecurity
The release reflects a broader shift in the AI industry.
Companies such as OpenAI and Anthropic primarily offer their most advanced models through hosted services, while Mistral is expanding its strategy around downloadable model weights.
Open-weight models give developers access to trained parameters, allowing them to run, customize and integrate models into their own systems.
This is particularly relevant to cybersecurity because organizations often handle sensitive information, including source code, network configurations, security logs and incident reports.
Keeping AI workloads within an organization’s own environment can help reduce dependence on external infrastructure.
However, open weights do not automatically mean that a model is completely open source. Access to trained parameters is different from unrestricted access to every part of the training data, development process and software stack. Organizations must also review the applicable licence before commercial deployment.
Mistral Strengthens Its Competition With Closed AI Labs
Mistral’s release comes as competition between open-weight and closed-model AI companies intensifies.
OpenAI, Anthropic and Google are investing heavily in advanced models, enterprise products and AI infrastructure.
Mistral is positioning Large 4 as an alternative for organizations that want advanced AI capabilities without depending entirely on external providers.
The company also claims that Large 4 uses 47% fewer tokens per task than GLM-5.3 on the Artificial Analysis Coding Agent Index comparison.
Lower token consumption can reduce inference costs in certain workloads, although actual deployment expenses will also depend on hardware, context length, throughput and system configuration.
For Mistral, the strategy is not simply to compete on model intelligence. It is also to compete on deployment flexibility, data control and operational independence.
The Bigger Picture
Mistral Large 4 highlights how the AI industry is moving beyond general-purpose chatbots toward specialized models that can perform complex work in areas such as software development and cybersecurity.
The release of downloadable weights gives enterprises and governments more control over how they use advanced AI, particularly when sensitive data and critical infrastructure are involved.
However, the same accessibility introduces security responsibilities. Organizations must protect their deployments against misuse, unauthorized access and vulnerabilities in AI-powered systems. The long-term success of open-weight cybersecurity models will depend on balancing capability, reliability and responsible deployment.
Looking Ahead
Mistral is likely to continue improving Large 4 through additional training, model customization and partnerships with cybersecurity organizations. Its reported benchmark performance could attract companies looking for AI-assisted vulnerability detection, code analysis and incident response without sending sensitive information to external services.
The larger challenge will be proving that these capabilities work reliably in real-world security environments. As downloadable AI models become more powerful, businesses will need to evaluate not only their performance but also their licensing, infrastructure requirements and safety controls before deploying them in critical systems.
Frequently Asked Questions
1. What is Mistral Large 4?
Mistral Large 4 is a trillion-parameter multimodal AI model developed by French AI company Mistral. It is designed for coding, reasoning, cybersecurity and other advanced tasks.
2. Why is Mistral Large 4 called Le Chonk?
Le Chonk is the informal nickname Mistral uses for its Large 4 model, reflecting its enormous one-trillion-parameter architecture.
3. Can businesses download Mistral Large 4?
Yes. Mistral has released the model weights through Hugging Face, allowing organizations to download and deploy the model on their own infrastructure, subject to its licensing terms.
4. How good is Mistral Large 4 at cybersecurity?
Mistral reports a 51% score on the Artificial Analysis Cyber Index and an 82% result on a real vulnerability reproduction and patching test. These are company-reported evaluation results.
5. Can Mistral Large 4 detect and fix software vulnerabilities?
The model is designed to assist with identifying vulnerabilities, reproducing security flaws and developing patches. Human security professionals still need to validate its findings and changes.
6. Is Mistral Large 4 open source?
Mistral describes the model as open-weight. This means its trained weights are available for download, but open weights are not necessarily equivalent to completely open-source software or unrestricted commercial use.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.
