Mistral Large 4: 1,000 billion parameters and open weights in October

What's new

Mistral has released Mistral Large 4 (internally ML4, nicknamed "Le Chonk") in public preview. It is the company's largest and most capable model to date. The preview API can be tried in Mistral Studio.

The model has 1 trillion parameters, or 1,000 billion. Parameters are the settings a model has learned during training, and roughly speaking, the more there are, the more the model can hold. The model is built as a mixture-of-experts: only a small portion is used at a time. Mistral reports 52 billion active parameters, while The Register puts it at about 49 billion. The model is natively multimodal, so it works with more than text, and it is trained on more than 160 languages, including all of the EU's official languages.

The weights, meaning the actual content of the trained model, have not been released yet. Mistral has set a date: they are due out by the end of October 2026, according to The Register on Hugging Face. Only then does the model become truly open.

What's clever

The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own data centers in Europe, and the preview runs on the same infrastructure. Mistral is planning a European rollout that the company will run itself from end to end, independent of other digital service providers and under European law. That could matter for companies and public authorities that want to keep data close.

Mistral highlights cybersecurity. On the Artificial Analysis Cyber Index, an independent test of how well models find and fix security vulnerabilities, ML4 ranks among the top five globally, according to Mistral. On a subtest in which the model must reproduce a real vulnerability in open source software and fix it, it scores 82 percent. That is Mistral's own run of the test. In Cybench, 40 exercises from security competitions, it solves 93 percent. Mistral writes that closed models such as Claude Opus 5.5 and GPT-6 Astra score close to zero on the same test. The explanation is cut off in the source.

Once the weights are out, organizations can self-host the model and run security work on their own terms, without the provider's refusals or the risk of losing access in the middle of an incident, Mistral emphasizes.

Cheaper or better?

The price has not been disclosed yet, either for the API or for regular users.

On quality, the picture is nuanced. Mistral says ML4 clearly outperforms all open-weight models from the US and Europe and is competitive with the strongest open models globally. That is the company's own claim. Independently, The Register places the model on Artificial Analysis' intelligence ranking between DeepSeek V4.1 Flash and OpenAI's GPT6 Luna. According to The Register, it is considerably better than Thinking Machines Labs' Inkling among downloadable, non-Chinese open models, but trails the closed flagships from OpenAI and Anthropic. Mistral's own tests show competitive performance against Chinese models from Alibaba, Moonshot, Z.AI and DeepSeek in coding, finance and law.

What it's good at

Mistral points to cybersecurity, finance and law as areas where the model is among the best open ones. In visual grounding, where the model must point to the right spot in an image, Mistral believes it beats even the top closed models.

The model was trained together with companies in finance, engineering, manufacturing, logistics, pharmaceuticals and the public sector. Until the weights are released, Mistral is testing a less moderated version with cybersecurity leaders, approved partners and government authorities. Mistral will publish more details on the architecture and more tests later, and the model is to form the foundation for a new generation of specialized models. The sources do not say whether or when it will be available in Denmark, but Mistral mentions European operation and multiple regions.