Mistral Large 4 ‘Le Chonk’ Is Now Public, With Open-Weight Claims

The company has released a public preview of its newest large language model, naming it Mistral Large 4, though some refer to it informally as “le Chonk,”. Mistral AI claims the model, which includes 1 trillion parameters and operates with 49 billion active parameters, is its largest and most capable offering yet.

Mistral Studio’s preview API is up and running right now, with full weights scheduled for release at the end of the month. The underlying system was designed with coding, cybersecurity, and multimodal comprehension in mind. According to Mistral, it already matches the top open-source models around the world, and beats every open-weight model produced in the US or Europe.

The Numbers Behind Le Chonk

The training for ML4 started from nothing across 3,800 NVIDIA Grace Blackwell GPUs housed in Mistral’s own European datacenters. A large portion of its training material covers more than 160 languages, with every official language of the European Union included among them.

The model scored 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. Combined, its Coding Agent Index score sits at 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max.

The Artificial Analysis Cyber Index ranks ML4 among the world’s top five models for cybersecurity, and it takes the lead over all other open-weight models outside China by a wide margin. When tested on its ability to reproduce a genuine vulnerability found in open-source software and then fix it, ML4 scored 82%, the highest mark of any model.

Why the Name Le Chonk Sticks

The nickname is playful, but the model it describes is anything but. Mistral says ML4 is built for “AI sovereignty” — the idea that countries and companies should control their own AI systems rather than rely on foreign providers.

Organizations can carry out their own security operations on their own servers using the model’s design, subject to their own policies. Mistral makes the case that provider-level refusals can prevent valid vulnerability research and incident response. The company also warns that losing access to a capability during an incident turns into a serious security hazard.

How ML4 Outperforms Closed Models

When it comes to how the system compares against rivals, the real point of interest lies in how several leading closed models respond to a specific vulnerability-reproduction test. Claude Opus 5.5 and GPT-6 Astra, for example, score near zero on that same test, simply because they refuse to carry it out.

The argument against refusing access rests on the fact that proving a flaw exists is exactly what defending software usually begins with — precisely the kind of work that safety filters in closed models can prevent. Attackers are now turning to those same models to aid their own offensive efforts, and security teams need tools that can keep up without facing the same denials.

Internal tests showed ML4 to be effective at analyzing malware, sorting out vulnerabilities by priority, and crafting detection rules. The model also handled 93% of the challenges in Cybench, a collection of 40 exercises taken from security contests — one of the highest scores recorded for an open-weight model.

The Road Ahead

The model will be made available across several regions around the world, according to Mistral, including a European deployment that the company runs entirely on its own, separate from other digital service providers and governed by European law. The company is also testing the model in real-world settings alongside cybersecurity leaders, vetted partners, and state authorities, who will use the same model with reduced moderation and enhanced cyber capabilities.

As the company moves toward releasing the weights, it has said it will share more information on the model’s design, its performance against other systems, and how it was fine-tuned after training. ML4 is being positioned as the base upon which a new family of tuned Mistral models will be built.

The preview API is currently up and running, and the firm is encouraging people to test it out and offer their thoughts online.

Key Facts Box

• Model: Mistral Large 4 (“le Chonk”)
• Parameters: 1 trillion total, 49 billion active
• Training hardware: 3,800 NVIDIA Grace Blackwell GPUs
• Coding scores: 61.7% DeepSWE v1.1, 59.4% SWE-Atlas-QnA, 28.3% Terminal-Bench 4
• Coding Index score: 49.8%
• Cybersecurity ranking: Top five globally, leads open-weight models outside China
• Vulnerability reproduction score: 82%
• Cybench score: 93%
• Weights release: End of this month

Mistral has built a system whose name “le Chonk” is a joke, while its goals carry genuine weight. It is wagering that open weights — models that anyone can run themselves — will answer to the demands of governments and companies that wish to hold sway over their own AI. The model’s showing on cybersecurity work backs up the seriousness of that ambition.

Source material: “Mistral Large 4: "Le Chonk",” Mistral AI.

AI · cybersecurity · LLM · Mistral · open-source