Mistral Large 4 Preview Launches with Weights Promised by End of October
Mistral released a public preview of its 1-trillion-parameter multimodal model Mistral Large 4 on October 6 2026. It leads open-weight models on cybersecurity benchmarks where…
- Preview API available now on Mistral Studio; weights scheduled for release by end of October 2026.
- Leads open-weight models on Cybench with 93 percent and an 82 percent score reproducing and patching a real vulnerability.
- Trained in Europe on 3,800 NVIDIA Grace Blackwell GPUs with multilingual data spanning over 160 languages.
What is available today
On October 6 2026 Mistral launched a public preview of Mistral Large 4, nicknamed le Chonk. The model is a natively multimodal mixture-of-experts system with one trillion total parameters and 52 billion active parameters.
It can be tried immediately through the preview API running on Mistral’s European infrastructure. The company states it delivers frontier performance on coding, agentic workflows, finance, legal tasks, science, and especially cybersecurity and visual grounding.
Weights are promised by the end of the month following further red-teaming with cybersecurity leaders, vetted partners, and state authorities. Those testers will use a version with reduced moderation and expanded cyber capabilities.
Cybersecurity performance
Mistral reports that on the Artificial Analysis Cyber Index, Mistral Large 4 ranks among the top five models globally and leads open-weight models developed outside China. It scores 82 percent on a test that requires reproducing a real vulnerability in open-source software and then patching it.
It also solves 93 percent of challenges in Cybench. Several leading closed models score near zero on the same test because they refuse the task. The source notes this matters because proving a flaw is often the first step in legitimate defensive work.
The model further shows usefulness in malware analysis, vulnerability prioritization, and writing detection rules. It is positioned for organizations that need sovereign, auditable AI able to run on private cloud or on-premise under their own policies.
Other benchmark results
On coding the model scores 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA, and 28.3 percent on Terminal-Bench 4.0. Its combined Coding Agent Index is 49.8 percent, ahead of several other leading models.
For agentic workflows it reaches 59.9 percent on AutomationBench and 1,393 Elo on AA-Briefcase. In multimodal tests it surpasses GPT-6-Astra on Dense 200 visual grounding. It also leads open-weight models on SciCode-Verified and outperforms others on finance and legal agent benchmarks according to third-party evaluators.
Safety evaluations show strong resistance to indirect prompt injections at 93.3 percent on Lakera’s B3 benchmark. However the source states its average refusal rate on malicious cybersecurity prompts from several jailbreak and harm benchmarks is higher than all other open-source models tested.
Why it matters for uncensored AI
Our analysis: the source shows that some closed models refuse vulnerability reproduction and patching tasks that defenders need, scoring near zero, while Mistral Large 4 performs strongly. Once weights are released, organizations can deploy the model on their own infrastructure and apply their own policies instead of depending on a provider’s moderation layer.
This gives security teams the documented choice of capability plus autonomy for legitimate work without the risk of sudden provider-level blocks during incidents. The higher cyber refusal rate noted in the source on malicious prompts means the preview still carries built-in safeguards; self-deployment after release will let operators decide how to handle that balance themselves.
