You can now download a leading AI model and run it on your own servers

The short version

• Mistral, the French AI company, has released its flagship model, Mistral Large 4, as a free download that companies can run on their own computers.
• On the tests Mistral reports, it is the strongest free-to-download model for writing code. It also does well at finding and fixing security flaws.
• For a bank or a government body in Indonesia, it is a serious option for AI that keeps sensitive data inside your own walls.

What Mistral released

A few weeks ago, Mistral opened Mistral Large 4 to the public as a preview. This week it released the model itself.

Most top AI models run only on their maker's servers, so your data has to leave the building. Mistral Large 4 can be downloaded and run on hardware you control.

The license lets you use the model, adapt it and build products on it. Only companies that resell the model as an online service need a paid license from Mistral.

If you would rather not run it yourself, you can use it through Mistral's own service or through major cloud providers.

What is new since the preview

Mistral spent three more weeks training the model after the preview. Most of the gains came in coding and security work.

• On a standard test of hands-on coding tasks, its score rose 13 points to 41%. That puts it ahead of every other free-to-download model.
• On an outside firm's security test, it scored 51%. The test checks whether a model can find a flaw in code and fix it without breaking anything else.
• It refused 96.6% of clearly malicious hacking requests. That is the highest rate Mistral measured among free models.
• In a test of attacks on AI assistants, such as tricking one into leaking its instructions, it resisted 97.1% of attempts. The preview resisted 93.3%.

Mistral published these figures and chose which tests to show. They are a good reason to try the model. Check it on your own work before you rely on it.

Mistral is also open about one risk. After watching how people used the preview, it says the model could give real help to attackers. It is tightening its safety limits and plans more updates in the coming months. A special version built for practice attacks on your own systems goes only to vetted partners, under contract.

Several European banks welcomed the release. Belfius and Crédit Agricole say the appeal is control: they can adapt the model and run it on their own systems.

What it means for companies in Indonesia

Many Indonesian firms want AI but hesitate to send customer data abroad. Regulated businesses such as banks and hospitals feel this most. A strong model that runs in your own data center, or with a local cloud provider, answers much of that concern.

It also changes how you pay. You pay for computers and the staff to run them, instead of a fee for each request. For heavy, steady use, that can work out cheaper. For light use, a hosted service is usually simpler.

There are costs to weigh. A model this large needs serious computing power. Your team also has to keep it updated and secure. Mistral trained it in Europe, so test it on Bahasa Indonesia and local documents first.

What to do next

• List the AI uses you have held back because of data rules. Those are your first candidates.
• Ask your IT team or cloud provider what hardware a model this size needs, and what it would cost each month.
• Run a small pilot on your own data in Bahasa Indonesia. Compare it with the hosted AI you use today.
• If security is the goal, start with the flaws in your own code. Keep a person approving every fix.
• Watch for Mistral's safety updates over the coming months.