Mistral released Large 4’s weights, touting cyber-defense results while warning its capabilities could also aid real attacks.
Mistral AI released downloadable weights for Mistral Large 4, a trillion-parameter model it says can help find and fix software vulnerabilities. The French company announced the release on October 29, allowing local deployment as well as hosted access. Its cybersecurity results are company-reported, and the model carries dual-use risks. 
TechCrunch covered the release.
Mistral Large 4 open-weight release
Mistral Large 4, also called “Le Chonk,” first appeared in public preview earlier this month. The newly released weights let developers download and run the model on their own infrastructure, rather than relying only on Mistral’s hosted API. Reuters reported that the release followed testing with government and corporate users.
The model card describes Large 4 as a multimodal mixture-of-experts system with one trillion total parameters. It activates about 49 billion parameters for each token, or 52 billion when embeddings and output layers are included. Mistral’s 
model card on Hugging Face gives those specifications and lists a one-million-token context window.
“Open-weight” means the model’s learned parameters are available to download. It does not necessarily mean the training data, training code or every part of the development process is public. Mistral’s model card says the weights use its Ecosystem Open-Weight License, a modified MIT-style license. It allows commercial and non-commercial use, but a separate commercial license is required to operate a model-as-a-service business.
Cybersecurity model claims and benchmarks
Mistral says Large 4 can help security teams identify vulnerabilities, reproduce them and develop fixes. In its 
release announcement, the company reported a 51% result on Artificial Analysis’ Cyber Index. The benchmark tests whether a model can find, validate and patch vulnerabilities in code, according to Mistral.
Those figures should be treated as benchmark results, not proof that the model improves every organization’s security in live use. Mistral says Large 4 refuses an average of 96.6% of malicious cyber requests in its tests. The company also acknowledged that its monitoring found potential for the model to assist real attacks, and said it was adjusting safeguards.
Mistral offers a more restricted cyber-specialized variant to selected enterprise design partners. The company says access requires vetting and contractual safeguards, with testing limited to systems the partners own or have permission to assess. That restricted variant is separate from the publicly downloadable general-purpose Large 4 weights.
What self-hosting can offer organizations
Local deployment can give an organization more control over where its model runs and how it handles internal data. Mistral says that can reduce reliance on outside providers during security work, including vulnerability research and incident response. Reuters reported the company described the release as a way for organizations to run the model independently.
“I’m not comfortable living in a world in which open weights do not exist.” — Pierre Stock, Vice President of Science at Mistral. He told Reuters that open access can protect users from losing access to a provider’s hosted service. However, the same access means users beyond carefully screened partners can obtain the general model weights.
Self-hosting also does not remove security responsibilities. Organizations must provide suitable computing infrastructure, manage access and monitor how the model is used. A model’s findings also need review by qualified security staff, since automated systems can produce incorrect or incomplete results.
Risks and limits of AI cyber tools
Cybersecurity tools can serve both defenders and attackers. A model that helps reproduce software flaws may support patching, but similar knowledge could help someone exploit systems. The public release makes that trade-off more significant because, unlike a hosted service, access to downloaded weights cannot be withdrawn from every user.
Mistral says its guardrails distinguish defensive work from malicious requests and that it will continue updating them. Reuters noted that the company acknowledges open access cannot be fully contained once weights are released. Benchmark refusal rates measure responses to specific test prompts, not the full range of ways a model might be adapted or used.
Ultimately, Mistral has made Large 4’s weights available for organizations to deploy themselves. Its cybersecurity and safety figures are company-reported, while its monitoring flagged possible attack risks. The specialized cyber version remains limited to vetted partners. Real-world use will test its value for defenders.
