Mistral Large 4 Takes on Defensive Cybersecurity

With 49 billion active parameters and an on-premises footprint, the French model targets defensive security teams blocked by commercial API filters.

French artificial intelligence lab Mistral AI released a public preview of Mistral Large 4 on Tuesday (October 6, 2026). The release brings a multimodal Mixture-of-Experts architecture totaling 1 trillion parameters, with 49 billion active parameters engaged per request. Mistral claims the system ranks among the strongest worldwide for defensive security tasks, pointing to independent tests focused on practical vulnerability analysis and incident response.

Security practitioners often deal with a recurring headache when working with major commercial models. Closed-source APIs frequently trigger automated guardrails whenever an engineer feeds them malware samples, exploit payloads, or reverse-engineering tasks. Because Mistral developed Large 4 specifically to support defensive teams, the model processes threat artifacts and unpacks malicious code without throwing false-positive refusal errors.

Mistral Large 4 in Defensive Benchmarks

Independent evaluations confirm strong capabilities on practical patching tasks. In the CyberGym-E2E-AA benchmark, run by Artificial Analysis to measure how well models replicate and fix real vulnerabilities in open-source projects, Mistral Large 4 scored 82%. That result edges past MiMo-V2.6-Pro at 79% and GPT-6 Luna at 78%.

Mistral also reported strong numbers on standardized industry evaluations. According to the company, Large 4 completed 93% of the 40 tasks in the Cybench suite, resolved 18 of 19 capture-the-flag challenges involving Cobalt Strike reverse engineering, and maintained 93.3% defense accuracy against indirect prompt injection attempts on Lakera’s B3 test.

Where the Scores Fall Behind

Leading in live vulnerability reproduction does not give Mistral an across-the-board victory. On the broader Artificial Analysis Cyber Index, which combines defensive tests with conceptual security reasoning, Large 4 scored 50 points.

That score earns the system a spot in the global top five and puts it third among open-weight options. However, it still trails the top tier: both Grok 4.7 and MiMo-V2.6-Pro lead that specific index with 56 points each. The French model stands out on hands-on code inspection rather than theoretical scoring.

European Infrastructure and Open Weights

Nicknamed “le Chonk” inside the company, the model was trained across Mistral’s European data center facilities using between 3,800 and 4,000 NVIDIA Grace Blackwell GPUs, as reported by CNBC and TechCrunch. That localized infrastructure setup reflects deliberate compliance planning, catering to European public sector bodies and enterprise security teams that cannot ship proprietary telemetry or incident data overseas.

For now, teams can test Large 4 through Mistral Studio and its hosted developer API. Full open weights are scheduled for free download in late October 2026, once external red-teaming reviews with enterprise partners and public regulators wrap up.