Mistral releases a trillion-parameter model anyone can download: whoever runs it owns the risk

Mistral AI, the French AI company, released the weights of Mistral Large 4 on 29 October, according to TechCrunch and Mistral's own announcement. The model has one trillion parameters and takes images as well as text. Mistral first put it in public preview earlier in October. Businesses and governments can now download it, run it on their own systems and fine-tune it. Mistral says it is the best open-weight model for coding, scoring 48% on the Artificial Analysis Coding Agent Index, effectively tied with the Chinese model GLM-5.3 while using 47% fewer tokens per task. Those are Mistral's own figures. The licence is a modified MIT licence, with a separate commercial licence required for companies that sell access to the model as a service.

Mistral is also promoting the model for cyber defence. It reports a 96.6% average refusal rate on malicious cyber requests, and offers a specialised cybersecurity variant to vetted partners. Its own page adds a caution: the company sees "a potential for uplift to conduct real attacks" and says it is modifying its guardrails accordingly. Microsoft says it is the first cloud provider to offer the model to customers.

Why it matters beyond the headline: with an open-weight model, the people who run it, not the supplier, decide what it connects to and who can switch it off. Mistral's Vice President of Science, Pierre Stock, told TechCrunch that this means access "can't be cut off by definition." That is the attraction. The same property means a supplier cannot withdraw the model if something goes wrong, and whoever hosts a copy can change how it behaves. For a school or an employer, that is the difference between buying a service and taking on a responsibility. Many will meet this model inside tools built by other companies, without being told.

Dan Fitzpatrick's analysis is that the open versus closed argument is being framed as a technology choice when, for leaders, it is a governance choice. Control and accountability arrive together. A school trust or a company that can run a frontier-scale model itself also has to decide who checks its safeguards, who monitors what it does and who answers when it fails. Software can help with each of those tasks. Only a named person can be responsible for them. The risk of waiting is real, because teachers and staff are already using AI ahead of policy, but adopting something nobody owns is the worse mistake.

What school leaders, educators, parents and business leaders should do now: ask every AI supplier which model sits underneath their product and whether it is open-weight or hosted by the supplier. Ask who maintains the safeguards and how changes are reported to you. Put one named person in charge of each AI tool in use, with the authority to switch it off. Before any tool built on a self-hosted model is used with pupils or customer data, check where the data goes. Then spend the time saved on teaching staff and young people to question what the model produces.

Dan Fitzpatrick works with schools, educators and organisations on the practical questions that follow a release like this: what to ask suppliers, who is accountable, and how to keep people doing the thinking.

Dan Fitzpatrick is available for interview and comment on this story. Available today for live or pre-recorded radio, TV, podcast and print interviews; broadcast-quality remote setup for live or pre-recorded down-the-line interviews. Contact: dan@theaieducator.io