Mistral 4: Europe's Biggest And Best Open-Weight AI Model Is Here.

Mistral Large 4 is here, and Europe just got its biggest AI model yet

Just a few days after Germany's Aleph Alpha surprised everyone with Kolibri, Europe has delivered another big moment. The French AI company Mistral AI has released Mistral Large 4, the largest and most capable model in its history. Inside the company it goes by the nickname "le Chonk", roughly "the big one", and after one look at the numbers you will understand why.

EUROPE JUST SHIPPED A 1-TRILLION-PARAMETER MODEL.

And here is the best part: you do not have to wait. You can chat with Mistral 4 on xPrivo from day one, privately, in the EU, and for free during the launch period.

What is Mistral Large 4 ("Le Chonk")?

Mistral Large 4 is a Mixture-of-Experts language model with about 1 trillion total parameters, of which roughly 49 billion are active per token. That means it has the knowledge capacity of a giant model while only using a fraction of its parameters for each answer. According to Mistral, it was trained on 3,800 NVIDIA Grace Blackwell GPUs.

Here is some key information at a glance:

• Developer: Mistral AI A European company from Paris, France.
• Nickname of Mistral 4 Large: "le Chonk"
• Size: about 1 trillion parameters in total, around 49 billion active.
• Training hardware: 3,800 NVIDIA Grace Blackwell GPUs (as stated by Mistral).
• Strong areas: cybersecurity tasks, agentic workflows, resistance to prompt injection.
• Availability: preview available now, weights planned for the end of October 2026.

It is Mistral's flagship, and it is a clear signal that European AI is not satisfied with playing in the small-model league.

What does "open-weight" mean, and why is it important?

Mistral plans to release the weights of Mistral Large 4 at the end of October 2026. Open-weight means that you can download the trained parameters of the model and run it on your own server or infrastructure. No black box, no single provider you are locked into and full privacy guaranteed.

This is a big deal for you or a company because leading AI providers in the US, such as Anthropic with Claude and OpenAI with GPT, do not allow their models to be run on users' own infrastructure. You would always need to trust the companies behind closed-source models, which do not release their model weights, not to read your sensitive prompts or train their next model on your chats. The benefits of open-weight releases are as follows:

• Sovereignty: the model can run on your own infrastructure in the EU under European rules.
• Privacy: hosting the weights yourself means your prompts do not have to leave your own environment.
• Independence: nobody can switch the model off or change it from one day to the next.

It is also the reason we are so excited at xPrivo. Open weights make it possible for us to bring Mistral 4 onto our own infrastructure in European once the model is out of the current preview. More on that in the rollout plan below.

Note: Mistral will first have security researchers, partners and government organizations test the model before the weights are published. We will update this article once the weights and the exact license terms are out.

How to chat with Mistral 4 for free on xPrivo

xPrivo is an open-source, privacy-first AI assistant from Europe that runs on European infrastructure by default. Using Mistral 4 takes three steps:

1. Open xPrivo Chat.
2. Open the model switcher and select Mistral 4.
3. Start chatting. It is free to use right now.

Where does your data go?

For now, Mistral 4 on xPrivo runs on Mistral's own infrastructure, located within the EU. That keeps your data in Europe and under European data sovereignty. Later, once the model is final and the weights are public, we plan to run it on our own infrastructure in order to maintain maximum control and ownership.

Our rollout plan for Mistral 4

Phase	What happens	Who gets access
Now (preview)	Mistral 4 is available on xPrivo, hosted by Mistral within the EU	Everyone, free for a limited time
Next	Free launch period ends	PRO users
Later (final release)	Model is out of preview and runs on xPrivo's own infrastructure	Default model for all xPrivo users

How good is Mistral 4? What we saw in our own tests

We did not want to simply repeat the press release, so we tested Mistral Large 4 ourselves. In our own tests it performed really well on logical reasoning questions and coding tasks. It also handled European languages better than the open-source models from outside the EU we compared it with. In our direct comparison it also came out ahead of Kimi K3.

That last point matters to us. European users need a model that handles German, French, Spanish, Italian and other European languages fluently, including their nuances and cultural context, rather than treating them as an afterthought. Many users and organizations also care about where a model comes from and which values shape it. Chinese models are well known for censorship: China's generative AI regulations require outputs to reflect "core socialist values", and models such as DeepSeek, Kimi, MiniMax, Qwen and GLM have been widely reported to avoid or deflect questions and criticism about historical and current Chinese events and political figures. A model built in Europe is developed within a European legal and cultural context, with expectations around transparency, neutrality and data protection. In our tests, Mistral 4 delivered on both counts: natural, accurate answers across European languages and balanced responses on sensitive political topics.

Mistral Large 4 benchmarks: what Mistral reports

Mistral compares its new flagship mainly with other open-weight models, including recent GLM versions from Z.ai and several DeepSeek releases. Here are the headline results as reported by Mistral:

Benchmark	What it tests	Mistral Large 4 result
CTF Speedrun	Security challenges under time pressure	18 of 19 tasks in 40 minutes (runner-up GLM 5.2: 16)
AutomationBench	Agentic task completion	59.9% (previous model: 6.3%)
B3 Agent Security Benchmark	Resistance to prompt injection and attacks	9 of 10 attacks resisted
DeepSWE 1.1 / Terminal Bench 4.0	Software engineering	Respectable among open-weight models

Source: Mistral's own announcement. Vendor benchmarks are a starting point, so test on your own tasks.

Where Mistral 4 shines

• Cybersecurity: the strong CTF Speedrun result shows how well it handles security tasks under pressure. Mistral also says the model does not refuse security-related work, which can help defenders, for example when analyzing new malware or rating vulnerabilities.
• Agentic tasks: the jump from 6.3% to 59.9% on AutomationBench is almost a tenfold improvement over its predecessor and puts it on par with open-weight competitors from China.
• Security against attacks: resisting nine out of ten prompt injection attempts is a solid result for a model that is meant to act as an agent.

Where we stay honest

Mistral's own charts show only open-weight competitors. In software engineering, the leading closed models from OpenAI and Anthropic score clearly higher, and they are not part of Mistral's comparison. Also, some details of the published results are missing, which makes a final judgment harder. Independent tests will show how well the numbers hold up in practice.

Mistral 4 vs. other open-weight giants

To put the size in perspective, here is how Mistral Large 4 compares with other large open-weight models, based on the parameter counts published for them:

Model	Provider	Parameters (total / active)
Mistral Large 4	Mistral AI (France)	~1T / ~49B
MiMo-V2.6-Pro	Xiaomi (China)	1T / 42B
GLM-5.3	Z.ai (China)	753B / 40B
Kimi K3	Moonshot AI (China)	2.8T / 104B
Kolibri	Aleph Alpha (Germany)	78B / 3B

Two European models, two very different philosophies: Kolibri is compact and efficient enough to run on two GPUs, while Mistral Large 4 goes big for complex tasks and would of course need much more GPUs. Europe now has both ends of the spectrum covered.

And don't forget Kolibri: also free on xPrivo

Only a few days ago, Aleph Alpha released Kolibri, a compact 78B model (about 3B active) under the Apache 2.0 license, strong in German and English. You can chat with Kolibri for free on xPrivo as well. Simply choose it in the model switcher.

Read our full article: Aleph Alpha Kolibri: Germany's New Open-Source AI Model

So right now, in a single privacy-first chat, you can switch between a German efficiency champion and a French heavyweight, both built in Europe.

A strong sign for Europe and a comeback in the AI race

Kolibri showed that Europe can build competitive efficient models from scratch. Mistral Large 4 shows that Europe can build at the very top of the open-weight scale. Together they send a clear message that European AI is slowly catching up in the AI race.

And it is open. Open weights mean more people can verify, host and improve what Europe builds, and it keeps control over the technology in European hands. At xPrivo, that is exactly what we stand for: privacy, transparency and European digital sovereignty.

Let's give this wonderful European model the attention it deserves. Try it, test it, push it with your hardest questions, and tell us what you think.

Mistral 4 FAQ

What is Mistral Large 4?

Mistral Large 4 is the largest model from French AI company Mistral AI, with about 1 trillion total parameters and roughly 49 billion active per token. Its internal nickname is "le Chonk".

What does "Le Chonk" mean?

"Le Chonk" is the playful nickname Mistral uses for Mistral Large 4. It roughly translates to "the big one" or "the chunk", a nod to the model's size.

Can I chat with Mistral 4 for free?

Yes. On xPrivo you can chat with Mistral 4 right away from day 1 for free for a limited time by selecting it in the model switcher. Afterwards it will be available for PRO users, and later it becomes the default model for all xPrivo users once it is final and running on our own infrastructure.

Is Mistral 4 private on xPrivo?

xPrivo is built privacy-first. For now, Mistral 4 runs on Mistral's infrastructure within the EU, which keeps your data in Europe. Our goal is to host it on our own infrastructure after the final release.

Is Mistral Large 4 open source?

It is planned as an open-weight model, meaning the trained parameters will be downloadable so you can host it yourself. Mistral plans to publish the weights at the end of October 2026, after a safety review. "Open-weight" is not the same as fully open source, and we will confirm the license terms once they are published.

Is there a Mistral Large 4 API?

Yes. Mistral offers the preview model through its own API. Once the weights are public, you will also be able to self-host it.

How does Mistral 4 compare with Kolibri?

Kolibri (Aleph Alpha) is a compact 78B model that is efficient and strong in German. Mistral Large 4 is a much larger 1T-parameter model aimed at complex tasks such as reasoning, coding and agentic workflows. You can try both for free on xPrivo.

Sources

• Mistral AI: Mistral Large 4 announcement https://mistral.ai/news/mistral-large-4/