Mistral Large 4: European giant model with a focus on cybersecurity

What's new

French AI company Mistral released its new model, Mistral Large 4, in public preview on October 6, 2026. It is available through Mistral Studio and Mistral's API and runs on Mistral's European infrastructure.

The model is multimodal, so it can take in both text and images. It is built as a Mixture-of-Experts model: it consists of many specialized "experts," and only some of them work on each task. In total it has about 1 trillion parameters, the numbers the model has learned. Mistral says 52 billion are active at a time, while Techzine and Runtime Wire report 49 billion.

The model was trained from scratch on 3,800 Nvidia Grace Blackwell GPUs in European data centers. It took about two months. According to Techzine, the final phase of reinforcement learning (training that rewards good answers) is still running.

What's smart

The context window is one million tokens. Roughly speaking, that corresponds to very long documents or large amounts of code that the model can keep track of at once. The training data covers 160 languages, including all of the EU's official languages. It supports coding, agentic workflows (where the AI solves tasks in multiple steps), function calling, structured outputs and questions about documents.

Mistral places particular emphasis on cybersecurity. According to Mistral, the model scores 82 percent on a test in which a known vulnerability must be reproduced and fixed, and 93 percent on Cybench. It is said to rank in the global top 5 on the Artificial Analysis Cyber Index.

The model is also intended to be open: the weights, meaning the model's internal numbers, are promised for release by the end of October, and several outlets cite October 27, 2026. That could let companies run it themselves and set their own safety limits. The license for the weights has not been disclosed, and the sources do not say whether the release has happened.

Cheaper or better?

In the API preview, Large 4 costs $0.70 per million input tokens and $2.10 per million output tokens. No monthly price for ordinary users has been announced. The sources also do not say how the prices compare with competitors'.

On quality, Mistral's own figures are all there is, and they have not been independently verified. Mistral reports 61.7 percent on DeepSWE v1.1, 59.4 percent on SWE-Atlas-QnA and 59.9 percent on AutomationBench. In a blind human evaluation of code quality, the model is said to rank second. On visual grounding (Dense 200), it beats GPT-6-Astra by 42 to 41 percent, according to Mistral.

Mistral also claims the model is the strongest outside China by a wide margin and is among the best open models in the world. The competitors in the coverage are Chinese models from DeepSeek, Alibaba, Moonshot and Z.ai (GLM-5.2), Reflection AI's Beam, and closed models from OpenAI and Anthropic. There are downsides too: Techzine notes that the model scores only 15 percent on Harvey's legal agent benchmark, which Techzine interprets as a sign that autonomous legal work is still hard for AI.

What it's good for

The model is aimed mainly at coding, workflow automation and security work. Mistral believes companies and governments can use it to defend against attackers who get closed models to break their own rules (jailbreaking). On the B3 AI Security Benchmark, it withstands 93.3 percent of attacks, according to Mistral.

For European companies, it may be an advantage that the model is trained and run in Europe and supports more EU languages. However, the sources do not explicitly say how availability looks in Denmark.

The rollout is staged: first a moderated API while Mistral continues red-team work and tests with select cybersecurity partners, then the weights. Runtime Wire warns that giving users control also makes it easier to remove safeguards once the weights are out. Independent tests and real-world deployments will have to show what that means in practice.