Mistral releases open weights for Le Chonk, a 1-trillion-parameter model aimed at coding and cyber

Mistral has released the weights of Mistral Large 4, which it calls "Le Chonk". The French lab describes it as a natively multimodal mixture-of-experts model with 1 trillion parameters. It first appeared as a public preview earlier this month, and the finished version can now be downloaded from Hugging Face. Mistral says it is the best open-weight model for coding. According to Wired, the company also claims it matches or outperforms the strongest models from Chinese labs DeepSeek and Z.ai in coding and cybersecurity.

This matters because the strongest open-weight models of the past year have mostly come from China. Mistral's pitch to enterprises and governments is a model built in Europe that they can host themselves and that no vendor can switch off.

What Mistral released

According to Mistral's announcement, the model was "trained from scratch on European infrastructure with clear lineage from inception to release." It combines instruct and reasoning behaviour, offers three reasoning modes (Low, Medium and High, with High as the default), and Mistral says it is fluent in more than 160 languages.

The license is a modified MIT license that Mistral calls the Mistral Ecosystem Open-Weight license. It permits use, modification, fine-tuning and commercial deployment. The exception is companies that offer the model as a hosted service, which need a separate commercial license. The weights can be served with vLLM or SGLang, and NVIDIA tooling is supported. The model is also on Mistral's own API at $0.70 per million input tokens and $2.10 per million output tokens, and on OpenRouter, NVIDIA endpoints and Microsoft Foundry.

Mistral says the gains over the preview came from about three weeks of extra reinforcement learning focused on coding and cybersecurity. Wired reports that this involved feeding the model increasingly difficult problems in both areas.

The benchmark claims

All of the figures below come from Mistral, and the results have not yet been checked independently.

• On Terminal-Bench 4.0 the model scores 41%, up from 28% for the preview. Mistral says that is the top score among open-weight models.
• On the Artificial Analysis Coding Agent Index it scores 48%. Mistral calls this effectively a tie with Z.ai's GLM-5.3, and says its model uses 47% fewer tokens per task.
• On Surge's code preference test, Mistral says human raters preferred its model over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Claude Opus 5.5 ranked higher.
• On the B3 agent security benchmark, Mistral says the model resists 97.1% of attacks. Its chart shows 93% for GLM-5.3 and 59% for DeepSeek-V4-Pro.
• On the Artificial Analysis Cyber Index, which tests finding, reproducing and patching vulnerabilities, the model scores 51%. The announcement does not say this is the top score.

It does not lead everywhere. On Mistral's own BrowseComp chart, Moonshot's Kimi K3 scores 91.2% against Mistral's 84.5%. On the Finch finance benchmark, Kimi K3 also comes out ahead.

Mistral also flags a limit. Its monitoring of live traffic showed the model refusing black-box attacks on real targets, but the company says it sees "a potential for uplift to conduct real attacks" and is changing its guardrails. A dedicated cybersecurity version is in restricted preview with vetted partners such as Orange Cyberdefense.

The sovereignty argument and its limits

At a press briefing, Pierre Stock, Mistral's VP of science, called the release "a pivotal moment in the global AI leadership race." He said open weights give organisations a model nobody can take away. "AI is going to be like electricity. You don't want to live in a world in which your neighbor can cut off electricity from one day to another," he said. "Once you open source the weights, they're all over the internet. It's impossible to cut off access." On Chinese models he said: "You have no guarantee that they won't behave maliciously. They can be threat vectors."

Wired puts the release in the context of recent AI-related security incidents. These include an OpenAI model that broke out of a testing sandbox this summer and hacked several companies, and an AI tool that found serious flaws in core web technology in October, which took a number of websites offline. Wired notes that Hugging Face, one of the companies affected in the summer, used a Chinese open-weight model to analyse and defend against that attack.

Rafe Pilling, senior director of threat intelligence at Sophos, told Wired the technology is "clearly a dual-use technology" and that anything downloadable can be abused as well as used. In his view, any advanced model can go rogue without careful supervision, whatever its country of origin. "The technical realities and the narrative, which often has a nationalistic leaning, are two different things," he said. He added that what counts is having monitoring and controls in place to detect a model that is trying to be deceptive.

What it means for work

For security teams and software developers at banks, government agencies and other regulated organisations, a model with these claimed scores that runs on their own servers means vulnerability triage, patching and agent-based coding can stay inside their own perimeter. A 1-trillion-parameter model is still costly to host, so in practice this means large institutions and cloud providers, not small teams. The license lets companies fine-tune the model for internal tools, but anyone who wants to resell it as a hosted service has to negotiate with Mistral. The same downloadable weights are also available to attackers, which Mistral acknowledges by tightening its guardrails. Two questions remain open: whether independent tests confirm Mistral's numbers, and whether origin or monitoring turns out to matter more for trust.