Mistral AI launched a public preview of Mistral Large 4 on October 6, 2026. It is a mixture-of-experts model with 1 trillion total parameters, 52 billion active, native multimodal input and a 1M-token context window. The weights are not yet downloadable: Mistral says they will be released by the end of October 2026. Until then, it is an API preview, not an open-weight model you can run yourself.
Mistral's announcement calls Large 4 its largest and most capable model and says it is still improving as Mistral refines it. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centers, and that the preview is served on the same infrastructure. Training data spans more than 160 languages, including every official EU language.
The Mistral changelog entry for October 6, 2026 lists the API identifier mistral-large-4, describes the architecture as mixture of experts with a 1M context window, and says launch pricing is 50% off for two weeks. Mistral's models overview lists it as version 26.10 with the license shown only as "Open," where earlier models such as Mistral Large 3 and Mistral Small 4 are listed as Apache 2.0.
Four things are not yet known or available:
Mistral also says that before releasing the weights it is red-teaming the model with cybersecurity leaders, vetted partners and state authorities, who get access with reduced moderation and expanded cyber capabilities.
Mistral's announcement leans hard on security work. It says Large 4 ranks among the top five models on the Artificial Analysis Cyber Index, scores 82% on one of that index's tests (reproducing and then patching a real vulnerability in open-source software), and solves 93% of the 40 Cybench challenges. It also states that Claude Opus 5.5 and GPT-6 Astra score near zero on the vulnerability-reproduction test because they refuse the task.
That contrast is the core of Mistral's argument: open weights let organizations run security research under their own policies, without provider-level refusals or the risk of losing access mid-incident. It is a real tradeoff, not a one-sided win. OpenAI, for example, describes GPT-6 Astra's stronger safeguards as a response to the model reaching the Critical cybersecurity level of its Preparedness Framework, in its GPT-6 Astra safety overview (see the GPT-6 Astra guide). Security teams will weigh capability and control against the risk that the same unrestricted capability is available to anyone once the weights are public.
All of these figures are reported by Mistral, including its characterization of competitor results, and this article does not reproduce them.
Mistral says Large 4 will be available in several regions, including a European deployment that Mistral operates end to end under European law, independently of other digital service providers. For European organizations with sovereignty requirements, that combination of a European-operated API now and self-hostable weights later is the main reason to watch this release.
European organizations with sovereignty requirements can test the preview API now and plan for self-hosting once weights ship.
Security teams frustrated by refusals in closed models are Mistral's explicit target audience. Test the preview on real workflows, and review your own policies for handling a less-restricted model.
Teams that need open weights today should look at releases that are already downloadable, such as DeepSeek-V4.1-Flash, whose model card states that the repository and weights are licensed under the MIT License (see the DeepSeek-V4.1-Flash guide), and revisit Large 4 when the weights and license are published.
When was Mistral Large 4 released?
A public preview launched on October 6, 2026. Mistral says the weights will be released by the end of October 2026.
Is Mistral Large 4 open source?
Not yet determinable. Mistral calls it open-weight, but the weights are not yet available and the license has not been named as of October 10, 2026.
How big is Mistral Large 4?
1 trillion total parameters with 52 billion active, per Mistral's announcement, with a 1M-token context window per its changelog.
Discover more content:
