Mistral has released the weights of Mistral Large 4, a 1-trillion-parameter open-weight model nicknamed “Le Chonk,” and says it is now the best open model for coding and defensive cybersecurity. Anyone can download it, run it on their own servers and adapt it. Here is what changed since the early preview and what it means for you.
What Mistral Large 4 Is
Mistral Large 4 is a natively multimodal model, which means it handles more than plain text. The French company first put a preview version out earlier this month. Mistral says three more weeks of reinforcement learning, a training method that rewards good answers on hard problems, produced large gains in coding and security work.
The company also says the model was trained from scratch on European infrastructure. That is a big part of its pitch for “sovereign AI,” where governments and banks want control over the technology they depend on.
Open-Weight Coding Results
Mistral reports that the new version leads open-weight models on two coding tests and ties the top model on a third. These are the company’s own numbers, so treat them as claims until outside testers repeat them.
| Test | Mistral Large 4 result | What it shows |
|---|---|---|
| FrontierCode 2.0 | 19.7%, first among open models | Whether a maintainer would merge the agent’s code change |
| Terminal-Bench 4.0 | 41%, up 13 points from preview | Coding work done in a terminal |
| Coding Agent Index | 48%, tied with GLM-5.3 | Agent coding, with 47% fewer tokens per task |
| BrowseComp | 84.5% | Finding answers by browsing the web |
In blind side-by-side tests run by Surge, human graders preferred its code over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked higher among the models tested.
Three Reasoning Modes
The model now offers three reasoning levels so you can trade speed and cost against accuracy. A light mode replaces the old “None” setting for simple tasks, and a top mode is meant for the hardest jobs that need careful reasoning.
Cybersecurity Focus and Safety Questions
Cybersecurity is the headline use case. Mistral says the model scores 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching software flaws. It also says the model refuses 96.6% of malicious cyber requests on average, the highest it has measured among open models.
Several security vendors and agencies, including Cisco, CrowdStrike and Europe’s cybersecurity agency ENISA, are quoted testing it. Mistral is also giving a specialized variant to vetted partners such as Orange Cyberdefense for penetration testing.
There is a catch, and Mistral admits it. After watching live traffic since the preview, the company sees “a potential for uplift to conduct real attacks” and is tightening its guardrails. Security researchers add that any model you can freely download can also be misused, since the weights cannot be switched off once they are public.
How to Try Mistral Large 4
You have several ways to try the model today:
- Mistral Studio and the Mistral API
- OpenRouter and Microsoft Foundry for a managed service
- Hugging Face weights to run yourself with vLLM or SGLang
- NVIDIA-hosted endpoints and NIM containers
API pricing is $0.70 per million input tokens and $2.10 per million output tokens. The license is a modified MIT license. It is free to use, change and sell with one condition: companies that offer the model as a hosted service must get a separate commercial license from Mistral.
What This Means for You
If you are a developer, a strong open-weight coding model gives you a cheaper option and the freedom to self-host. You can compare it against other open-weight models on your own code before you commit.
If you run a company with strict data rules, the European training story and private hosting may matter more than a benchmark. If you work in security, it is worth testing on threat detection tasks, while keeping in mind the dual-use risk.
- Run your own tests before trusting vendor benchmarks.
- Check the license if you plan to resell access.
- Review your security guardrails if you host any open model.
The release also shows the open-model race is not only about China versus the US anymore. Europe wants a seat at that table, and similar open model releases keep raising the bar. Mistral has been busy elsewhere too, including its forecasting model.
Sources
Primary source: Mistral’s Large 4 announcement. Additional reporting from WIRED.
Conclusion
Mistral Large 4 is now a downloadable, trillion-parameter model that its maker says leads open models in coding and defensive security. The numbers come from Mistral and its partners, so independent testing will matter. For now it is a serious new option for developers and organizations that want control over their AI, with real safety trade-offs to manage.
