With Le Chonk, Mistral AI Has the Most Powerful AI Model Outside the US and China

French AI company Mistral AI has claimed a notable place in the global AI race with Mistral Large 4. Nicknamed Le Chonk, the model earned the highest overall score of any AI model developed outside the United States and China, according to independent benchmarking firm Artificial Analysis. But its individual test results tell a more nuanced story.

Mistral unveiled Large 4 on October 6, 2026. AI Wereld previously covered its public preview and the planned release of its model weights. The question now is how competitive France’s model really is against its international rivals.

An independent analysis by Artificial Analysis gives Mistral Large 4 a score of 38 on the Intelligence Index. This composite benchmark measures how language models perform across a range of tasks, including reasoning, coding and processing complex information.

That marks a major milestone for Mistral. According to this benchmark, the French company has built the most capable model from a developer outside the two dominant AI nations.

Large 4’s score of 38 is roughly level with GPT-6 Luna in the tested configuration. DeepSeek V4.1 Flash scores slightly higher, at 39. Several other American and Chinese models also rank above Large 4.

So Mistral has not built the world’s smartest model. But it has shown that Europe can train a competitive AI model on its own.

The picture shifts when you look at specific applications. A general intelligence score measures different capabilities from a specialist test focused on, for example, software security.

Mistral’s technical announcement puts a strong emphasis on cybersecurity, financial analysis and complex business processes.

Large 4 scores 50 on Artificial Analysis’s Cyber Index, putting it level with GLM-5.3-Flash and ahead of several other well-known models.

On one specific security test, Mistral scores around 82 percent. The test assesses whether a model can reproduce existing software vulnerabilities and then fix them.

That matters to security specialists, who increasingly use AI to inspect source code, analyze potential vulnerabilities and suggest fixes.

But the results need context. Some closed models refuse certain cybersecurity tasks because of their safety safeguards. As a result, a model can score lower without being less capable at coding.

Mistral also needs to guard against these capabilities being used for attacks. The company is therefore having selected security specialists and government agencies conduct additional tests before releasing the model weights.

Mistral Large 4 contains around one trillion parameters—the internal values the AI model learns during training.

But the system does not use all of them at once.

Mistral uses a so-called Mixture-of-Experts architecture. At each processing step, the model activates only a selection of its specialized components. Depending on how they are counted, around 49 to 52 billion parameters are active.

This approach helps a large model run more efficiently. It does not mean Large 4 can run on an ordinary laptop. The full model still requires substantial memory and specialized hardware.

Mistral trained the model using around 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers.

That points to a significant strategic shift. Europe does not have to rely solely on American AI models; it can also manage the infrastructure that runs advanced models.

That gives Europe more control over availability, data processing and how AI services are built. At the same time, the European AI sector remains dependent on international suppliers for components such as chips.

Mistral plans to make Large 4’s weights available by the end of October. Until then, developers can use the model through its public API preview.

The distinction matters. Downloadable model weights let organizations run an AI model on their own infrastructure and potentially adapt it to specific tasks, depending on the terms of the license.

For banks, hospitals, government agencies and industrial companies, that control can be decisive. They may not want to depend on an external provider to access sensitive documents or support business-critical processes.

Cost is another factor. Using standard rates, Artificial Analysis calculated a cost of around $1.13 per completed benchmark task. Comparable models from DeepSeek and other providers were significantly cheaper in this test.

So a competitive price per million tokens does not automatically make a model the cheapest option for completing a task.

For European companies, the trade-offs are becoming more concrete: what does a model deliver, what does it cost in practice, and how much control do they retain?

Mistral Large 4 matters most because it expands that choice. The real test comes when outside developers can download the promised model weights and independently reproduce its results.