Mistral AI has put Mistral Large 4 into public preview, but the one thing that makes it “open-weight” is still missing: the downloadable weights. For now the model runs only on Mistral’s own servers, reachable through a preview API on its Mistral Studio platform. The Paris-based lab says the weights will be released by the end of the month, and its launch post gives no exact day.
The size comes with two numbers, and the second one matters more. Mistral says the model holds 1 trillion parameters in total but uses 49 billion of them for any given token. Picture a huge library where each question sends the model to a few shelves: all the books must be stored, but only a small fraction gets read per answer. That gap is what makes a model this large affordable to serve. Mistral has not yet published architecture details, so how it achieves the split remains unstated.
Open-weight means you can download the model and run it on hardware you control, instead of calling someone else’s service. Mistral’s pitch leans on exactly that for security teams. The post also does not name the licence, which decides whether a company can use the weights commercially. That detail is absent from the announcement.
The cybersecurity claims are the boldest. Mistral says the model ranks among the top five on the Artificial Analysis Cyber Index, an outside evaluation of finding and fixing flaws in real software. On the index’s reproduce-and-patch test it reports 82 percent, which it calls the highest of any model, and it attributes the near-zero marks for Claude Opus 5.5 and GPT-6 Astra on that same test to refusals: those models decline the task. Read that gap carefully. It measures willingness as much as skill, and defenders are the buyers Mistral is courting.
Coding and agent results come from Mistral’s own runs. It reports a Coding Agent Index score of 49.8 percent, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max, and 59.9 percent on AutomationBench, a set of 657 business workflows. A blind rating by annotators at Surge AI placed it second of five models at 3.74, behind Claude Opus 5 at 4.22. On visual grounding, it says it edges GPT-6 Astra on Dense 200, 42 percent to 41 percent. A one-point margin on a self-chosen test is not a ranking.
The geopolitical framing is Mistral’s own. The post presents Large 4 as a European model trained on 3,800 Nvidia Grace Blackwell GPUs in the company’s own datacenters, with a European deployment it operates under European law. It also describes the model as outperforming any open-weight model built in the US or Europe, while claiming only parity with the strongest open models worldwide. That is a sales position, and the announcement includes no independent head-to-head against those models.
Two versions are also in play. Mistral says vetted cybersecurity partners and state authorities are red-teaming a separate build that has looser moderation and a broader set of cyber abilities. The public preview is the standard one, so the refusal-free numbers may not describe what ordinary users can call today. The post also nicknames the model “le Chonk,” a name it uses alongside ML4.
Mistral ties the release to its €3 billion Series D, which Mistral calls the biggest equity raise any European tech company has completed, and says the training run is still in progress. That means the preview is a moving target, and scores could shift before the weights ship.
Security and platform teams weighing a self-hosted model should hold off on any procurement decision until the weights, the licence terms, and independent reruns of the cyber tests all exist, since today only Mistral’s numbers are on the table.
Mistral AI, “Introducing Mistral Large 4,” company blog, undated post, with secondary coverage dating the announcement to 6 October 2026.
