Mistral previews Le Chonk with unusually strong cyber benchmarks

What Mistral announced

Mistral AI announced Mistral Large 4 on October 6 and opened a public API preview in Mistral Studio. The model’s nickname, “Le Chonk,” fits its scale: Mistral says it contains one trillion total parameters, with 49 billion active for each token. It is natively multimodal, so text and images are handled inside one model rather than through a separate image system.

The company says the model supports more than 160 languages and was trained on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own European data centers. That infrastructure detail is unusual in a launch post, but it does not reveal the training corpus, data mixture or full evaluation methodology. The release is still a preview, and Mistral says training and refinement continue.

How the one-trillion-parameter design works

Mistral Large 4 is a mixture-of-experts model. Instead of using every parameter for every token, a routing mechanism selects a smaller subset of specialized expert blocks. The claimed 49-billion-active figure therefore matters more for per-token computation than the one-trillion headline, while the full count describes learned capacity distributed across experts.

That design can improve the trade-off between capacity and serving cost, but it creates operational questions around memory, routing and hardware. Our mixture-of-experts explainer explains why active parameters and total parameters should not be compared as if they were the same metric. Mistral has not yet published enough deployment detail to estimate realistic self-hosting requirements.

Access now—and what open weights will mean later

Developers can currently access the model through Mistral Studio as an API preview. Mistral says the weights will be released by the end of October, so teams should not plan a self-hosted deployment as if those files already exist. The preview may also change before release because training and refinement are ongoing.

A downloadable checkpoint is not automatically open source. The license can restrict commercial use, redistribution or deployment, and the hardware bill may be substantial. Review open weights versus open source before treating access to parameters as access to the full training recipe.

Teams should also apply an AI model license checklist once the final terms appear, rather than assuming the preview’s API terms predict the weight license.

Why the cyber numbers are notable

Mistral reports that Large 4 ranks among the top five models on the Artificial Analysis Cyber Index. It also claims an 82% success rate on an internal test that asks the model to reproduce and patch real vulnerabilities, plus 93% on Cybench. If the tasks and scoring hold up, those results point to serious capability for security analysis, code review and controlled red-team work.

The company separately reports 93.3% resistance on Lakera’s B3 adversarial benchmark, a KORA score of 1.691 out of 2, and the highest average cyber-refusal rate among the open-source systems in its comparison. Those are safety or robustness signals, not evidence that the model lacks dangerous capability. A system can be highly capable at cyber tasks and still resist some malicious instructions; the two axes should not be collapsed into one secure-model score.

All figures come from Mistral’s announcement. Until prompts, sampling rules and independent replications are available, they should be read as vendor-reported results. Our AI benchmark guide explains why evaluation harnesses and dataset selection can materially change a headline.

What builders can use it for

Mistral also reports software-engineering and agent results including 61.7 on DeepSWE, 59.4 on SWE Atlas, 28.3 on Terminal Bench and 49.8 on its coding-agent evaluation. The company lists 59.9 on AutomationBench and an Artificial Analysis Briefcase Elo of 1393. These are relevant to codebase work, tool-using agents and long, multi-step enterprise tasks.

Practical evaluation should begin with representative internal work. A security team might test triage and patch suggestions inside a sandbox; a software team might measure repository-level fixes with human review; a multilingual organization could test the languages it actually serves. API-preview behavior, latency, pricing, data terms and refusal patterns matter as much as benchmark position.

Limitations and the next checkpoint

Mistral’s post is primary evidence for the architecture, access plan and reported scores, but it is not an independent audit. The company has not yet released weights, a full technical report, detailed cyber artifacts or reproducible serving guidance. Top-five status also depends on which competitors and benchmark version were included.

Strong vulnerability reproduction may help defenders, but it can also lower the skill required for misuse. High refusal scores do not guarantee that an attacker cannot bypass safeguards, combine benign-looking steps or exploit a downstream tool. Deployments with repository access, terminals or networks need least-privilege permissions, sandboxing, logging and human approval.

The end-of-month weight release is the next concrete checkpoint. Watch for the final license, model card, checkpoint sizes, quantized variants, context details, inference recipes and safety documentation. Le Chonk is a consequential preview, but independent testing will decide how much weight its capability and robustness claims deserve.