Mistral Large 4 'Le Chonk' is a 1-trillion-parameter open model built for cyber defense

The size: a 1-trillion-parameter, natively multimodal model, with 49 billion parameters active at a time. That's a mixture-of-experts design, where only part of the model works on each task.

Text and images: it can process both, and Mistral calls it a "step change" in its models' ability to understand images.

The claim: Mistral says it's its largest and most capable model to date, competitive with the strongest open models in the world, and well ahead of any open-weight model developed in the US or Europe. Mistral co-founder and chief scientist Guillaume Lample described it as "a new generation of models" at a press conference, Euronews reports.

Languages: a significant share of the training data covered more than 160 languages, including every official language of the European Union.

Now, via API: developers can try a preview through Mistral's API on Mistral Studio.

Weights later this month: Mistral says it will release the weights by the end of the month. The Next Web and Euronews put the date at October 27. Open weights mean anyone can download the model and run it on their own servers.

Until then: Mistral is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities. They get a version with reduced moderation and expanded cyber capabilities.

Mistral says it trained Large 4 from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own data centers in Europe, the same infrastructure that now serves the preview. Other outlets, including The Next Web and Euronews, report about 4,000 GPUs over two months, and The Next Web says training used about 10 megawatts of power.

Customers can run it on their own servers or use Mistral's API in a region of their choice, including a European deployment that Mistral says it operates end to end, independent of other digital service providers and under European law.

Mistral says companies and governments shouldn't depend on a provider that could switch off their tools at any moment, The Next Web reports.

Security is the center of Mistral's message.

The scores: on the Artificial Analysis Cyber Index, an independent evaluation of how well models find and fix security flaws, Mistral says Large 4 ranks in the top five globally and leads open-weight models developed outside China by a wide margin.

On one of the index's tests, which asks a model to reproduce a real vulnerability in open-source software and then patch it, Large 4 scored 82%, the highest of any model, Mistral says. It also solves 93% of Cybench, a set of 40 security competition exercises.

The refusal argument: Mistral says several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on that same test because they refuse the task. "Yet defending software often starts with proving that a flaw is real," Mistral writes, and closed models' safety filters can block exactly that.

"The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyber attacks," Lample said, according to The Next Web.

The safety side: Mistral says Large 4 refuses malicious cyber requests more often than all other open models on average, based on prompts from JailbreakBench, StrongREJECT and AgentHarm. On Lakera's B3 AI Security Benchmark, it resists 93.3% of prompt injection attacks, the company says.

Mistral published a long list of results. All are preliminary, and many are its own.

Coding: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4, using Artificial Analysis numbers. In a blind human evaluation with Surge AI, Large 4 ranked second of five models (3.74 out of 5), behind only Claude Opus 5 (4.22).

Agents: 59.9% on AutomationBench, 657 business workflows across apps like Gmail, Google Sheets, Slack and Salesforce, ahead of Kimi K3 and DeepSeek V4 Pro.

Finance and law: Mistral says third-party evaluator vals.ai found it beats GPT-6 Astra on representative legal and financial tasks. On Harvey's Legal Agent benchmark, it scored 15%, ahead of Kimi K3 at 13% and GPT-6 Astra at 5%, The Next Web reports, though the low scores show how far all models are from handling legal work alone.

Vision: on the DIOR-RSVG remote-sensing test, it scored 73% versus 68% for GPT-6 Astra, according to The Next Web. On Dense200, Mistral says it edges GPT-6 Astra 42% to 41%, while The Next Web reports both at 42%.

Real-world uses: Mistral wants to put the vision skills to work helping insurers assess storm damage in aerial photos, utilities inspect power lines and farms monitor crop health, The Next Web reports. It can also turn technical drawings into CAD models, and Mistral says it performs well on semiconductor tests.

Science: Mistral says Large 4 is state of the art among open-weight models on SciCode-Verified, a test of implementing scientific workflows in code, and can generate a full Hartree-Fock chemistry simulation in one shot.

Training at scale: at its current scale of about 3,000 GPUs, Mistral says a single reinforcement learning run produces roughly 33 billion tokens a day.

Still training: the reinforcement learning run behind the preview is still in progress, and Mistral says the model shows "no signs of saturation." Results are expected to change before the weights come out, The Next Web reports.

Unverified: there are no independent benchmarks based on the open weights yet. Mistral also says Large 4 beats top open models from Kimi, DeepSeek and Meta on cyber tasks without sharing the figures, according to The Next Web.

Crowded field: the launch came a day after US startup Reflection unveiled its first open-weight model, Beam, whose performance claims also haven't been independently verified, Euronews notes.

Mistral says Large 4 is the first milestone funded by its €3 billion Series D, which it calls the largest equity round ever raised by a European technology company. Euronews reports the round valued Mistral at more than €21 billion.

The open-weight race has taken on a geopolitical edge, with Chinese developers setting the pace and many US companies keeping their best models closed, Euronews notes. Pierre Stock, Mistral's first employee and vice president of science, said the model is "basically stronger than China's models from this summer" in some respects.

Stock compared building a new AI model to "building a spaceship," and argued that US companies use reports of rogue AI agents to "market rogue events" as a reason why "only a few trusted players can take care of the technology," Euronews reports.

• If you're a developer: you can try Large 4 in Mistral's API today, and download the weights later this month.
• If you work in security: Mistral is offering a strong model with fewer refusals that you can run on your own hardware, once the weights ship.
• Everyone else: check independent benchmarks after the weights release before trusting the headline numbers.