Mistral releases Le Chonk, a huge open AI model from Europe

Le Chonk is a playful nickname for Mistral Large 4, and it is all about size. “Chonk” is internet slang for something big and chunky, usually a well-fed cat, and Mistral bolted a French “Le” on the front. The official name is Mistral Large 4, sometimes shortened to ML4.

The size is real. Parameters are the internal settings a model adjusts while it learns, and Large 4 has roughly 1.05 trillion of them. It uses a design called mixture of experts, where only a small slice of the model wakes up for each word it produces. In this case that is about 49 billion parameters, or about 4.7% of the total, which keeps running costs closer to a mid-sized model.

The catch is memory. Even if only a sliver is active, the whole model still has to sit in a computer’s memory, so this is data center software, not something for a laptop.

Cyber defense is the headline job. Mistral designed Large 4 to read through system logs, audit code for weak spots and check for newly published security holes. It can also probe systems the way an attacker would, so defenders can patch problems first.

Mistral reports scores of 93% on Cybench and 82% on CyberGym-E2E, two tests of hacking and security skill. It also points out that several closed models from rival labs score close to zero on the second test, mainly because they refuse to reproduce a vulnerability at all, even though proving a bug is real is routine defensive work.

The second big focus is agents, meaning AI that plans and carries out long jobs on its own, such as research or coding tasks that would take a skilled person hours or days. Pierre Stock, Mistral’s vice president of science, has called agentic ability the core skill the team was chasing.

There are smaller specialties too:

• Images: Large 4 reads pictures as well as text, and Mistral says it beats rivals at analyzing aerial and satellite photos, for example spotting forest fires on a map.
• Languages: its training covered more than 160 languages, with all of the EU’s official tongues among them.
• Long documents: it can take in up to 1 million tokens (chunks of text) at once, enough for very long reports or codebases.
• Niche work: Mistral tuned it for finance, manufacturing and electrical engineering tasks that bigger labs tend to overlook.

Mistral calls Large 4 the most capable open-weight model built outside China, and “very, very close” to some closed ones. Outside testing gives a more mixed picture. In a blind review run with Surge AI, professional raters gave the preview 3.74 out of 5, second of five models tested, ahead of China’s GLM-5.3 and Kimi K3 but behind Anthropic’s Claude Opus 5 at 4.22.

Early third-party results also show it matching or slightly beating DeepSeek on financial tasks and keeping pace with pricier closed models on some image tests. Final benchmark numbers had not been published at launch, and some coding scores were run privately, so they can’t yet be checked by outsiders.

The resources gap is striking. Mistral says it trained Large 4 from scratch in its own European data centers over two months on Nvidia Grace Blackwell chips. The company’s statement put the figure at 4,000 chips, while the model documentation lists 3,800. Either way it is a small fraction of the roughly 100,000 chips Nvidia’s Jensen Huang said OpenAI used for GPT-6 Astra. Mistral also stresses that it did not train on the answers of rival models, a shortcut some Chinese labs have been accused of.

Money is less of a worry than it was. In September Mistral raised $3.3 billion at a $24 billion valuation, the largest funding round ever for a European tech company.

The timing is no accident. In June the Trump administration put temporary limits on how models from OpenAI and Anthropic could be distributed, citing fears they could be used for serious cyberattacks. For any company that relies on a closed model to guard its network, that was a reminder that access can vanish.

Guillaume Lample, Mistral’s co-founder and chief scientist, argues the point applies to American firms too: what counts is owning the model, because a closed one might not be there tomorrow. Customers will be able to run Large 4 on their own hardware or through Mistral’s cloud, including a European region where data stays under EU law.

For everyday people the change is mostly behind the scenes. You probably won’t chat with Le Chonk directly, but banks, software firms and public agencies could use it inside their own systems, which means more AI tools that keep your data in-house rather than sending it to an outside provider.

Price helps the pitch. The preview API costs $1.36 per million input tokens and $4.18 per million output tokens, with cached input at $0.14 per million.

The download is the part still pending. Mistral’s launch page promised the weights by the end of October, along with architecture details, more benchmarks and notes on how the model was trained. A wire report pointed to Oct. 27, but Mistral itself never confirmed a day. As of Oct. 26, the model’s Hugging Face page showed an upcoming-release countdown to Oct. 31 at 15:00 UTC, with a note that the date could still change.

Mistral has spent the gap testing the model with security leaders, vetted partners and state authorities, who get a version with fewer restrictions and stronger cyber skills.

The bigger unknown is the license. Mistral Large 3 came under the permissive Apache 2.0 license, but the Large 4 license file was still missing in late October. Industry reports expect a custom license instead, and Mistral’s current Modified MIT terms already require companies earning more than $20 million a month to sign a commercial deal. If you plan to build on Large 4, read that license before you commit any servers to it.