French artificial intelligence company Mistral has publicly released the weights of its Large 4 model, allowing organisations to download and operate the system on their own infrastructure while targeting cybersecurity teams that need greater control over sensitive data and defensive tools.
The model, known internally as ML4 and nicknamed “Le Chonk”, was first unveiled in preview earlier this month.
Mistral says the newly released version has undergone additional training and testing with government agencies and corporate users, improving its coding, reasoning and cybersecurity capabilities before the public release.
Unlike AI services that primarily operate through provider-controlled online platforms, organisations can deploy Large 4 on their own servers and determine how the model is integrated into their security environments.
Mistral Pitches Self-Hosting For Cybersecurity Teams
The ability to run Large 4 locally is central to Mistral’s cybersecurity pitch.
Security teams frequently work with highly sensitive information, including proprietary source code, vulnerability reports, system configurations and incident-response data.
Sending such material to an external AI service may create privacy, regulatory or operational concerns.
An open-weight model gives organisations the option to keep those workloads inside their own infrastructure.
Mistral argues that this can also reduce dependence on external AI providers during security incidents.
If a third-party service restricts access to certain cybersecurity tasks, experiences an outage or changes its usage policies, a locally deployed model can continue operating under the organisation’s own rules.
That autonomy is particularly useful for activities such as vulnerability research, malware analysis and incident response, where security teams may need models to examine potentially dangerous code or perform actions that general-purpose AI services sometimes restrict.
Large 4 Improved After Three Weeks Of Testing
Mistral released a preview of Large 4 on October 6 before making the model weights publicly available.
The company says it spent the intervening weeks conducting additional reinforcement learning and gathering feedback from customers, government organisations and security partners.
The updated model includes new reasoning modes and improvements aimed at making it more efficient and customisable.
Large 4 contains about one trillion parameters, with a smaller number activated during individual tasks, and supports multimodal inputs.
Mistral describes it as its most capable open model so far for areas including coding, cybersecurity, finance and legal work.
The company says performance improved significantly during the period between the preview and the public release, particularly on coding and agentic benchmarks.
As with other vendor-published benchmark claims, those figures provide an indication of performance but may not predict how the model performs across every real-world security environment.
Open Weights Bring Capability And Risk
Making powerful AI models downloadable provides users with greater control, but it also creates an obvious security trade-off.
Once model weights are publicly available, the developer can no longer decide who uses the system or how it is modified.
Organisations can customise the model, remove safeguards and operate it privately.
That freedom benefits researchers and defenders who want independent access to advanced capabilities.
It can also potentially benefit malicious users seeking to adapt the same model for offensive cybersecurity work.
Mistral has acknowledged that tension.
Company executives argue that open-weight AI is important because it prevents advanced technology from becoming concentrated entirely in the hands of a small number of providers.
At the same time, once an open-weight model is released, access cannot realistically be withdrawn.
That makes responsible deployment and local security controls especially important.
Cybersecurity Becomes A Major AI Battleground
Large AI models are increasingly being evaluated not only on general reasoning and coding but also on their ability to find vulnerabilities, analyse attacks and generate security fixes.
Google, OpenAI, Anthropic and other AI developers have been testing models on vulnerability discovery and exploit-related benchmarks.
Open-weight models from Chinese developers have also shown increasingly strong cybersecurity capabilities.
Mistral is positioning Large 4 as a European alternative in that landscape.
The company says the model can help organisations identify vulnerabilities and fix weaknesses while allowing them to retain direct control over the model and the data it processes.
The model was trained using Nvidia Grace Blackwell GPUs in Mistral-operated European data centres, reinforcing the company’s broader argument around European AI infrastructure and technological sovereignty.
Orange Cyberdefense Moves To Integrate The Model
The cybersecurity positioning has already attracted enterprise interest.
Orange Cyberdefense, the digital security arm of French telecom group Orange, announced a partnership with Mistral to integrate a cybersecurity-focused version of the company’s latest AI model into its products.
The partnership gives one of Europe’s larger cybersecurity providers access to a European-developed model capable of being adapted for security operations.
The agreement comes as security companies increasingly explore AI for threat detection, incident analysis, vulnerability research and automated response.
For defenders, the attraction is speed.
An AI model capable of understanding code, analysing technical evidence and interacting with security tools could reduce the time required to investigate incidents or identify vulnerable systems.
But organisations adopting such models will still need controls governing what systems they can access, what actions they can perform and how their outputs are validated.
Large 4 therefore represents both a technological release and a broader experiment in how powerful cybersecurity AI can be distributed.
Instead of keeping advanced capabilities behind a provider-controlled interface, Mistral is betting that organisations will increasingly want to own, customise and operate those capabilities themselves.
