What happened
On 29 October Mistral published the weights of Mistral Large 4 on Hugging Face, three weeks after making the model available through its API as a public preview. It is a mixture-of-experts model with roughly a trillion parameters, 52 billion of which are active for each token. Mistral says it was trained from scratch on its own infrastructure in Europe. The company positions it mainly for coding and cyber defence.
The licence is called the Mistral Ecosystem Open-Weight License, a modified MIT licence. It allows commercial and non-commercial use with one exception: a company running a Model-as-a-Service business needs a separate commercial licence from Mistral. These are open weights, not open source in the full sense.
You can download the model and run it yourself (vLLM, SGLang). The examples on the model card assume an eight-GPU server, and a separate quantised variant targets NVIDIA Blackwell cards. If you would rather not own hardware, Mistral's API charges USD 0.70 per million input tokens and USD 2.10 per million output tokens. The model is also available through OpenRouter, NVIDIA and Microsoft Foundry, including a private deployment on Azure.
The benchmark results, including the claim that this is the best open model for coding, come from Mistral. At the time of writing there are no independent measurements.
Our read
This is our interpretation. In June we wrote that cloud sovereignty is becoming a procurement criterion: the question of where your AI stack runs, and under whose jurisdiction, is moving into the tender scoring table. Until now that row had one weakness. The strongest models were closed and reachable only through APIs from vendors outside the EU, so "the European option" usually meant a compromise on quality. Large 4 doesn't fully fix that, but it narrows the gap. For the first time, a model of this scale from a European vendor can be downloaded and run on your own infrastructure, with nobody's permission needed for each request.
Open weights do not mean "free, in-house". A trillion parameters is a cluster decision, not an API key. It means a node with eight Blackwell-class cards, power, cooling, and people to run and update it. For most companies the realistic route to this model is still the API or a cloud deployment, and then the jurisdiction question comes back in another form: whose cloud, which region, who can see the logs.
The second point is the licence. The Model-as-a-Service exception is short and its boundary is not obvious. An internal assistant for employees fits ordinary commercial use. A product where customers pay for access to the model's answers may already count as a model service. Nothing we read shows exactly where Mistral draws that line. That is a question for a lawyer and for Mistral, not for the development team.
The third point is the benchmarks. The numbers come from the vendor, and the model is compared mainly with other open models. We expect, and this is speculation, that on Polish documents and in narrow processes the gap to the closed models you already use will be larger than the tables suggest. Only your own data can settle that.
Why it matters
Private Equity
If a portfolio company sells an AI-based product, ask which licence its model runs under and whether the way customers use it might count as a model service under that licence. Open weights reduce the risk of being cut off by a vendor, but they don't remove licence risk. In due diligence that is one more row, and a concrete one.
Enterprise
For banks, energy, healthcare and public administration, where data cannot leave your own infrastructure, this is the first European candidate of this scale worth a real test. Before a hardware budget appears, though, separate two requirements that often get merged: "data cannot leave our servers" and "data must stay under EU jurisdiction". The first needs your own cluster; the second may be met by a European deployment over an API. Once that distinction is drawn in enterprise projects, it usually turns out that only some processes need on-premises.
SMB / mid-market
Don't buy graphics cards. If a European vendor matters to your customers or in tenders, test the model through the API on one task and compare the result with what you use today. The input price is low, but calculate cost per task, not per token.
One move for this week
Take one process that is stuck because "data cannot go outside" and write down what exactly is blocking it: a regulation, a customer contract, internal policy, or habit. For each blocker, mark whether it requires on-premises or whether EU jurisdiction is enough. Then run 50 historical cases from that process through Large 4 and through the model you already use. The hardware conversation only makes sense once the result is good enough and the blocker really requires your own server room. Describe your case: mailto:[email protected]?subject=Rozmowa%20z%20Aurora%20AI