Mistral has put the files of its newest flagship model, Mistral Large 4, on Hugging Face, a few weeks after opening it as a preview. Anyone can download it, change it and use it in their own business for free. The one exception is a company that sells access to the model as a service, which needs a separate licence from Mistral.
That is a bigger deal for ordinary businesses than the benchmark scores suggest, and the reason has little to do with Mistral.
A rented model has a lifespan set by someone else
Every AI model you reach through a company's API is on a calendar you do not control. Anthropic's schedule retires Claude Sonnet 4.5 on November 30, 2026, a little over a year after it launched. OpenAI took GPT-4o out of ChatGPT in February 2026, and its API follows eleven months later. Rented models have been living for about a year.
Picture a hospital group's finance team that taught a model to read supplier invoices and match them to purchase orders. Getting it right took weeks: example invoices, a list of odd cases, a check by an accountant before anything was paid. When the model is retired, the team starts that work again on the replacement, because a newer model reads the same invoice slightly differently. For a team with a dozen workflows like this, that is probably two to four weeks of skilled time every year, spent standing still.
A file you downloaded has no retirement date. The invoice reader behaves the same way in 2028 as in 2026, for as long as you keep a server running it.
What owning the file does and does not give you
It gives you the right to stay. It does not give you a reason to, and it does not make the model cheap to run. This model is a trillion-parameter system, and its smallest version is a 610 GB download that needs a server with four to eight of NVIDIA's top data-centre GPUs. A hospital group can afford that; a shop or a school department cannot, and should not try.
For them the benefit is indirect but real. Because the weights are public, the model is also available through Microsoft's Foundry, through Mistral's own API at $0.70 per million input tokens and $2.10 per million output tokens, and through other hosts. If one of them drops it, the file moves to another. That is a very different position from a closed model, where the only place to buy it is the company that can stop selling it.
The licence matters here. Mistral's previous flagship, Large 3, came under the looser Apache 2.0 licence. Large 4 keeps the freedom to use it inside your own business, but a company that wants to resell it needs Mistral's permission, so there will be fewer competing hosts than there would have been.
Pin the work that never changes, rent the work that does
An open model is a floor, not the best available. Anthropic's Claude Opus 5.5 was preferred in 78% of blind code comparisons against Mistral Large 4. So the sensible split is by kind of work. High-volume, stable jobs such as sorting invoices, tagging support tickets or reading forms can run on a model pinned as a file, where nothing changes under you. Hard, changing work, where the newest model matters, stays rented. I expect most companies that use AI heavily to end up doing both within a year or two.
Security teams are an early case. They can run the model on their own logs and code without sending either outside, and, as we covered when it launched as a preview, it takes on security work rivals refuse. The same openness has a cost: a downloaded model's guardrails are in the hands of whoever runs it. Treat it like any server software and decide who in your company may run it, and where.
Before you build a workflow on any AI model, ask what happens to it on the day its maker retires it. For a rented model the answer is that you start again. For one you hold as a file, the answer is that you decide.