Mistral Large 4: large European model with a focus on cyber defense

What's new

Mistral AI released Mistral Large 4, nicknamed "Le Chonk," in public preview on October 6, 2026. It can be tried through Mistral Studio and Mistral's API. The model understands both text and images and covers more than 160 languages, including all of the EU's official languages.

Under the hood, it is a so-called mixture-of-experts model. That means the model is built from many smaller "experts," and only some of them work on each task. In total it has about 1 trillion parameters (the model's adjustable knobs). Mistral states 52 billion are active at a time, while The Register reports roughly 49 billion. It was trained on about 3,800 Nvidia Grace Blackwell GPUs in European data centers.

Mistral has said the model's weights (the trained model itself) will be released as open weight by the end of October. Several outlets cited October 27 as the target, but as of October 31, the model card on Hugging Face showed the weights were not yet available, and the license had not been published. It is therefore unclear how open the license will be.

What's smart

Mistral puts the emphasis on cybersecurity. According to Mistral, Large 4 scores 82 percent on a test of reproducing vulnerabilities and 93 percent on Cybench. The company says the model is among the top five globally on Artificial Analysis' Cyber Index and leads among open-weight models from outside China. According to AI Magazine, Artificial Analysis confirms 81.7 percent on the vulnerability test.

Mistral also claims that several closed models, including Claude Opus 5.5 and GPT-6 Astra, score close to zero on the cyber tests because they refuse the tasks. Mistral says Large 4 does not refuse legitimate cybersecurity tasks. That is Mistral's own assessment.

On coding, Mistral reports 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA and 28.3 percent on Terminal-Bench 4.0. Mistral also says training was still running at the time of the announcement and that the model shows no signs of saturation.

Cheaper or better?

Mistral's announcement lists a price of $0.70 per million input tokens and $2.10 per million output tokens (tokens are small chunks of text). Other sites cite a list price of $1.36 and $4.18, so the price may have changed, or the figures may come from different providers. A monthly price for regular users has not been disclosed.

On quality, the picture is mixed. In a blind human evaluation by Surge AI, the model ranked No. 2 of 5 for code quality with 3.74 out of 5. Claude Opus 5 came in ahead, and according to AI Magazine it scored 4.22. According to The Register, an independent analysis places Large 4 between DeepSeek V4.1 Flash and GPT-6 Luna, and it performs better than Thinking Machines' open-weight model Inkling. Artificial Analysis gives it 38 on its Intelligence Index and 49.8 percent on its Coding Agent Index, ahead of DeepSeek V4 Pro and Qwen3.8 Max.

Mistral says the model beats GPT-6 Astra on finance and legal tests and on visual grounding (42 versus 41). These are Mistral's own figures, and only some have been independently confirmed.

What it's good for

The model points mainly toward cyber defense, where companies can use it to investigate and understand vulnerabilities. It is also intended for coding, automation (59.9 percent on AutomationBench) and work in finance and law. On the B3 test, it withstands 93.3 percent of attacks.

For European companies, it is notable that Mistral emphasizes European infrastructure and sovereignty. The training was funded by a Series D round of EUR 3 billion. The sources do not make clear who can get access in Denmark and the EU, but the model is available through Mistral's API and Studio. According to The Register, it can run on 8-way GPU setups such as Nvidia HGX B300 or AMD MI355X, so self-hosting requires heavy-duty hardware.