Mistral opens weights for its trillion-parameter Large 4 model

Mistral AI released the open weights for Mistral Large 4, which it calls ML4, on 29 October 2026. According to TechCrunch, the company made the weights available through Hugging Face and its own API, so that other companies and government agencies can download the system and deploy it themselves. The model had been in public preview on Mistral's API since 6 October 2026, and the full release followed three more weeks of training.

What was released

Mistral describes Large 4 as a mixture-of-experts model. In plain terms, it holds about one trillion parameters but activates only a subset for each task. Mistral's figures, as reported by The Next Web, put the active portion at 52 billion parameters per task. The model reads text and images and produces text and code. Mistral is also releasing a quantised version, a compressed form of the model that its release notes say runs on four Nvidia B200 or B300 graphics processors.

The cyber figures need care. Mistral says the model scores 79% on CyberGym-E2E, a test that asks whether a model can find, reproduce and patch real memory-safety bugs in open-source C and C++ projects. Heise reports that the figure places the model level with Xiaomi's MiMo-V2.6-Pro. Artificial Analysis, an independent evaluator, measured the earlier preview at 81.7% on its own version of the test. The two numbers come from different versions of the model and of the test, so the 79% should be read as the company's own figure for the released weights.

Why it matters for businesses that automate work

Until now, the most capable models of this scale were available mainly through a provider's API. Open weights change where the data goes. A company that runs the model on its own servers can keep documents, customer records and security logs inside its own network. Mistral's own release says this is the point: organisations can inspect, customise and run the model on their own systems.

The trade-off is cost and staffing. A model this size needs a serious hardware budget and people who can run it, patch it and monitor it. For most firms, the API or a managed cloud service will be cheaper to start with. The open weights matter most for organisations that must keep data in-house for legal, contractual or security reasons.

What changes in practice for a business that automates work

For a business already using AI for routine work, such as drafting reports, sorting invoices or writing follow-up emails, the release adds a choice. The same workflow can run on a hosted API, on a cloud provider that offers the model, or on hardware the company controls. The workflow itself does not change. What changes is who holds the data and who pays for the compute.

Three practical points stand out:

• The licence is a contract. Mistral's licence allows free use, modification, fine-tuning and commercial use. It requires a separate commercial licence from Mistral for any business that operates the model as a service for others. Firms that only use the model internally are not in that group, but the licence text should be read by legal counsel before deployment.
• Hardware sets the floor. Mistral's quantised version targets four high-end Nvidia GPUs. Few small firms have that in-house, so most will start with a cloud or API route and move to their own servers only if data rules demand it.
• Model updates become your job. A self-hosted copy stays the same until you change it. Security patches, safety settings and performance changes must be tested and approved by your own team.

A model that can help find and patch software flaws is also a model that needs human review. Cyber output should go to a qualified person before any system is changed. Results on a company's own code and systems may differ from the published benchmarks, and only testing on that environment will show how well the model performs.

What to watch next

Three things will show whether the release changes buying decisions. The first is independent testing. Artificial Analysis and similar groups have already published results for the preview, and their figures for the released weights will matter more than the company's own charts. The second is licence uptake. Cloud providers and resellers will need a commercial agreement with Mistral before offering the model as a paid service, and the first such deals will show how the clause works in practice. The third is Mistral's stated plan to release more models from the same family in the coming months, as reported by SiliconANGLE.

Businesses do not need to act on this release today. The useful step is to list which routine tasks involve sensitive data, and to decide which of them could ever run on hardware the company controls. That list will be the starting point for any evaluation.

This article reports the release and its published terms in general terms. It is not legal advice on the licence for your case.