Blog
Mistral opens a public preview of Mistral Large 4, with open weights promised by month-end
On 6 October 2026 Mistral launched a public API preview of Mistral Large 4, a 1-trillion-parameter multimodal model with 52 billion active parameters, trained in its own European data centres. Mistral says the weights will be released by the end of October. Its benchmark claims are its own and have not yet been checked independently.
On 6 October 2026 Mistral launched a public preview of Mistral Large 4 (ML4). You can use it through the preview API on Mistral Studio now. Mistral says the weights will be released by the end of this month, together with more details on the architecture, more benchmarks and its post-training method.
The model in brief
From Mistral's announcement:
- Size: 1 trillion total parameters, 52 billion active. Mistral describes it as a mixture-of-experts model that combines instruction-following and reasoning in one model.
- Inputs: natively multimodal, so it can work with images as well as text.
- Training: trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European data centres. The preview is served on the same infrastructure.
- Languages: Mistral says a large share of the training data was multilingual, covering more than 160 languages, including every official EU language.
- Deployment: it will be available in several regions, including a European deployment that Mistral runs end to end under European law.
Why the weights are delayed
Mistral says that until the weights come out, it is red-teaming the model in real-world settings with cybersecurity leaders, vetted partners and state authorities. Those testers get the same model with reduced moderation and expanded cyber capabilities.
The cybersecurity pitch
Security is the headline use case. Mistral argues that provider-level refusals in closed models can block legitimate vulnerability research and incident response. It says open weights and self-hosting let organisations run security work under their own policies.
Mistral reports that ML4:
- ranks in the top five globally on the Artificial Analysis Cyber Index;
- scores 82% on one of that index's tests, which asks a model to reproduce a real vulnerability and then patch it. Mistral says this is the highest score of any model, and that some closed models score near zero because they refuse the task;
- solves 93% of Cybench challenges.
Mistral also says ML4 refuses malicious cyber prompts more often than the other open models it compared.
Coding and agents
Mistral's reported scores include 61.7% on DeepSWE v1.1, 28.3% on Terminal-Bench 4 and 59.9% on AutomationBench, a set of 657 business workflows across apps like Gmail, Sheets, Slack and Salesforce. In a blind human evaluation of coding quality run with Surge AI, the preview placed second of five models, behind Claude Opus 5.
How much to trust the numbers
These are launch figures chosen and published by Mistral, for a preview it says is still improving because its reinforcement-learning run is "still in flight". Until the weights are out and other people have tested them, treat the figures as claims.
The bigger picture
Mistral calls ML4 the first milestone on a roadmap funded by its €3 billion Series D. It plans to use ML4 as the base for a new generation of specialised models.
What this means for you
For European organisations, or anyone who wants to run a strong model on their own hardware, ML4 is worth watching. Try the preview API now, but wait for the weights release and independent tests before planning production work around it.