--- title: "Mistral Large 4 is in preview, with open weights due this month" url: https://blog.vomini.com/blog/mistral-large-4-preview-open-weights date: 2026-10-07T13:30:00+00:00 updated: 2026-10-27T22:42:33.292790+00:00 author: "Roberto R." section: "News" language: en site: "Vomini" license: free to quote with attribution --- # Mistral Large 4 is in preview, with open weights due this month > Mistral opened a public preview of Mistral Large 4, a 1 trillion parameter model trained in its own European data centers, and says the weights will be published by the end of October. Mistral pitches it for security… On 6 October, Mistral opened a public preview of Mistral Large 4 through its Studio API, and said it will release the model's weights by the end of the month. Open weights means anyone can download the model and run it on their own servers, instead of only calling it through Mistral's service. Mistral calls it its largest and most capable model so far. It has 1 trillion parameters in total, with 52 billion active for each token, a design where only part of the model works on each piece of text, which keeps it cheaper to run than its size suggests. It reads images as well as text. ## Built and served in Europe Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own data centers in Europe, and that the preview runs on the same machines. It will be available in several regions, including a European deployment that Mistral says it operates end to end under European law, without depending on other digital service providers. More than 160 languages were part of the training data, including every official EU language. ## The pitch on security Mistral leans hardest on cybersecurity. It says the model ranks among the top five on the Artificial Analysis Cyber Index, and scores 82% on one of its tests, which asks a model to reproduce a real vulnerability in open source software and then patch it. Mistral says that is the highest score of any model, and that Claude Opus 5.5 and GPT-6 Astra score near zero on that test because they refuse the task. Mistral argues that defenders often have to prove a flaw is real before they can fix it. Until the weights are out, Mistral says it is testing the model with security leaders, vetted partners and state authorities, who get a version with reduced moderation. It also says the public model refuses malicious cyber prompts more often than other open models on three public test sets. ## Other numbers Mistral reports 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4 for coding, and 59.9% on AutomationBench for business workflows. These are Mistral's figures, mostly compared with other open models. The post does not list an API price for the preview. ## What to do this week If your company cannot send data to US cloud providers, or needs to run models on its own hardware, this is the one to put on the shortlist. Try the preview in Mistral Studio on a few of your real documents now, so you are ready to judge whether running the open weights yourself is worth the hardware when they arrive. ## What we think A frontier-sized open model trained in Europe gives companies there a real alternative to sending everything to a US provider. The security argument cuts both ways: once the weights are public, the same capability to reproduce vulnerabilities is available to anyone who downloads them, and only the model's own refusals stand in the way. ## Sources - [Mistral AI](https://mistral.ai/news/mistral-large-4/), 6 Oct 2026 --- Originally published at https://blog.vomini.com/blog/mistral-large-4-preview-open-weights