Mistral Large 4 opened on October 6 as a public preview API, and the weights, which are the only reason anyone calls it open, arrive at the end of the month. CNET puts the rollout at October 27. Until then, it is a model you rent from Mistral, and everything I know about how good it is comes from Mistral.

The numbers on the box: one trillion parameters, natively multimodal, 49 billion active at a time, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own data centers in Europe. That last clause is the sovereignty pitch in one line, and it lands harder than it did in August, when I wrote that Large 3 ran DeepSeek’s blueprint and Mistral had started hosting Chinese weights outright. This time the company is saying it trained the thing itself, on its own hardware, under European law. The silicon is still American, so what it can claim is who operates the machines, and I’d keep that distinction in mind every time the word sovereign shows up in the coverage.

Mistral says the model scores 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4, and that its Coding Agent Index of 49.8% beats DeepSeek V4 Pro 0813 and Qwen3.8 Max. It also says the model beats every open-weight model built in the US or Europe, while CNET’s version of the same briefing says it is catching up to the Chinese labs. Those two framings don’t quite agree, and I haven’t seen an independent run of any of it. None can happen at scale until the weights are out.

The cybersecurity section is where Mistral spends the most effort, and where I have the most mixed feelings. It reports 82% on an Artificial Analysis test that asks a model to reproduce a real vulnerability in open-source software and then patch it, which it calls the highest of any model, plus 93% on Cybench. The sales point is that Claude Opus 5.5 and GPT-6 Astra score near zero on that same test because they refuse. I understand the argument; a defender who needs to prove a flaw is real gets nothing from a model that won’t touch the exploit. But the preview is also being red-teamed by cybersecurity leaders, vetted partners, and state authorities using a version with, in Mistral’s words, reduced moderation and expanded cyber capabilities. So there are already two tiers of access before the weights exist, and once they are public, the moderation is whatever each downloader decides to keep. I haven’t seen a safety card for any of this, and I’d like one before the 27th.

The money behind it is the record €3 billion round with Samsung’s biggest check from September, and this is the first flagship since. European buyers are watching for exactly that kind of proof, because Lagarde’s warning about Europe being cut off from AI turned a policy worry into a procurement question, and banks like BNP Paribas are choosing carefully what moves to a US cloud and what stays put. A local model that holds up against the American and Chinese ones gives them something to point at. A local model that is competitive on Mistral’s slides gives them a preview.

I’ll look again when the weights land. If someone else’s harness reproduces the 82%, this is a big deal; if it doesn’t, the cyber pitch was a marketing choice.

Sources