Mistral Large 4 gives you more choice over where AI runs

Start with the job you want done

If you are choosing AI for an app, start with one piece of work you can check. It might be turning an approved document into a clear answer. Keep that example when you compare tools, so you can judge the result against something you understand.

On October 29, Mistral released an improved version of Large 4 that organisations can download and run independently. That follows an online preview earlier in October. The downloadable files, often called weights, hold what the model learned during training.

The change gives technical teams another option for building AI into a useful product. Mistral also offers access through its own online service, so you can explore the model without taking on the whole job of running it yourself.

Read the announcement: Mistral: Large 4 downloadable release announcement.

More control brings a running job

Running a model on infrastructure you manage gives you more say over its setup and operation. That can matter when deciding where business information goes and how a tool fits into existing work.

The full download is about a terabyte, roughly 1,000 gigabytes. Running it requires specialist hardware and technical setup. An ordinary laptop is not a sensible starting point for this full model.

Ask who will operate the service, keep it updated and help when it stops working. Include those responsibilities in the price comparison. Access to the files alone does not tell you the total running cost or how well the model will handle your task.

Check the terms for your plan

Mistral distributes Large 4 under its Ecosystem Open-Weight License. It permits commercial and non-commercial use, with extra conditions for businesses offering access to models as a service.

The licence sets out exceptions, including internal use and model capabilities embedded in specific product features. Ask your builder to read the terms for your planned service and check whether a separate commercial agreement is required. Being able to download a model does not settle every use question.

Make the choice easier to judge

If you are learning, use made-up or approved information for a first comparison. If you are commissioning a tool, ask for the same exercise before agreeing to a change.

1. Write one task and what a good result must include.
2. Try the same approved example with your current tool and the proposed model.
3. Check accuracy, missing information and how easy the answer is to use.
4. Compare the full running cost, information handling and support plan.

Build understanding alongside the app

Our perspective is that more model choices are useful when you can explain why one fits the work. You can practise that judgement now, even if a technical partner handles installation and hosting. Bring your example and the result you need to a training or planning session.