Article — "Mistral Ships 1T-Parameter Open Weights Trained on European GPUs": Last week, the open-weight cost shock was a question: can models released outside the US and China actually hold prices down? This week, Mistral gave the question a physical address. On October 29, the Paris lab published the full weights of Mistral Large 4 — a 1-trillion-parameter, natively multimodal Mixture-of-Experts model, reportedly 49 billion active parameters per token (though some sources say 52 billion), 128 routed experts, and a context window reported as 1 million tokens by the model card and Neural Brew, though Singularity.Kiwi reports 512K — under a license that is MIT with one deliberate hole punched in it. But before the license, the part that matters most is the part you can measure in watts.
The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe, reportedly drawing roughly 10 megawatts over about two months, per preview-era reporting. The reinforcement-learning phase reportedly ran on Mistral's own compute in France, according to The Next Web. That is the sentence to sit with. A European lab took a fixed quantity of electricity, a fixed count of accelerators, and a fixed block of calendar time, and reportedly produced a frontier-class open-weight model without renting a single hour from a US hyperscaler, though the evidence pack does not address whether any US hyperscaler compute was used.
Arthur Mensch's framing, as quoted by The Next Web, is the whole argument in one line: "Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion." Whatever you think of the benchmark gap to DeepSeek and GLM, that sentence is the actual news. Frontier training is starting to look less like a capital moat and more like an engineering problem with a known budget.
Frontier training is starting to look less like a capital moat and more like an engineering problem with a known budget.
Then you read the license, and the sovereignty story gets complicated in an instructive way.
Mistral released the weights under something it calls the Mistral Ecosystem Open-Weight License — a modified MIT that permits commercial and non-commercial use, modification, fine-tuning, and commercialization. With one carve-out: operating a Model-as-a-Service business requires a separate commercial license negotiated with Mistral. Internal use is unaffected. Running it through Mistral's own products is unaffected. Building a business whose product is access to this model requires a conversation.
That is not a loophole. It is the architecture of the entire release.
Think about what sovereignty-by-download was supposed to mean. A government, a bank, a hospital network downloads the weights, runs them inside its own perimeter, and stops depending on anyone's continued goodwill. Pierre Stock, Mistral's VP of Science, put the rationale plainly: AI is going to be like electricity, and no company should depend on someone who can cut off its access "from one day to another."
That logic holds for self-hosting. It does not hold for the layer most of the economy actually consumes AI through.
If a European cloud provider wants to offer Le Chonk to its customers — and that is precisely the layer where sovereign alternatives to US hyperscalers would need to live — it must go through Mistral. The weights are free; the right to resell them is not. Sovereignty-by-download turns out to be a hedge against dependency at the infrastructure layer, paired with a toll booth preserved at the services layer.
The defense against this reading is that it is exactly what open weights have always been commercially. Llama had its acceptable-use policy and its 700-million-user threshold. The difference is that Mistral is selling sovereignty as the product, and a toll booth on the resale layer is a strange thing to put inside a sovereignty pitch.
So what does Europe actually get?
Three things, and they are not nothing.
First, a procurement option that does not route through US hyperscaler goodwill. Mistral operates a European deployment end-to-end, independently of other digital service providers and under European law — and the model is natively fluent in all official EU languages, trained with a significant multilingual share spanning 160-plus languages. For public-sector buyers whose rules increasingly favor jurisdiction-bound suppliers, that is a checkbox no American lab can currently tick.
Second, a cyber-defense capability that closed models structurally refuse to provide. This is the sharpest part of the release. On CyberGym-E2E — finding, reproducing, and patching real memory-safety bugs in open-source C and C++ — Large 4 reportedly scores 79%, per The Next Web. Mistral claims 51% on the Artificial Analysis Cyber Index, though preview-era independent numbers collected by I-SCOOP had it at 50, level with GLM-5.3-Flash and behind Xiaomi's MiMo-V2.6-Pro at 56, with an independent re-run on public weights still pending. The company's own framing of the gap is brutal: per Mistral, several leading closed models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on a vulnerability-reproduction test because they refuse the task. Defending software often begins with proving a flaw is real — exactly the work safety filters can block. As Guillaume Lample put it at the October 6 preview press conference, enterprises and governments need to defend against threat actors who are jailbreaking closed models to attack them.
Third, a proof point. Mistral is not claiming to have beaten DeepSeek on the aggregate Intelligence Index — it scores 26, the highest among Western open-weight models but fifth among open weights overall, per The Next Web and Neural Brew, and the model card says it is the highest among western open-weight models with leading Chinese open models scoring higher. What it is claiming is that the resource ceiling for reaching this class of capability is far lower than the capital-markets narrative suggests. That claim is hard to test, per The Next Web, given how little the Chinese labs disclose about their training costs. But the physical receipts here are unusually legible: 3,800 GPUs, 10 megawatts, two months.
The efficiency angle deserves its own beat, because it is where the 10/19 question gets a partial answer. On the Artificial Analysis Coding Agent Index, Large 4 scores 48% — effectively tied with GLM-5.3 — while using 47% fewer tokens per task on average. Terminal-Bench 4.0 rose from 28.3% at preview to 41% at release, with output tokens down 30%. Stock argues that per-token pricing misleads when a model needs far fewer tokens per job, and on this evidence he is right. The pressure that open frontier weights exert on closed-lab pricing may not come primarily through undercutting on price. It may come through making the unit of work cheaper.
Here is the uncomfortable question the release leaves open.
If open weights at the frontier are the goal — if the point is that no one can switch off your capability, that defenders can work without asking permission, that a small country can own rather than rent — then why does the license preserve exactly the commercial structure open weights were supposed to disrupt?
The charitable answer is that Mistral is a company with a €3 billion Series D and, per Euronews, a valuation above €21 billion, and it needs a revenue layer that survives its own giveaways. The less charitable answer is that "open" is now a spectrum, and the frontier of that spectrum is being defined by whoever has the weights to define it.
Neither answer changes what happened in those European datacenters. Europe now has a frontier-class open-weight model it trained itself, on its own soil, on its own power budget. That is a real capability, and it is the first time the sentence has been true.
It is also a capability with a meter attached — not on the electricity, but on the resale. Sovereignty-by-download works right up to the point where you want to sell it to someone else.