Mistral Large 4 Goes Open Weight, With a Hosting Catch

Mistral has released the weights of Mistral Large 4, the model it nicknames “Le Chonk”, three weeks after opening it as an API-only preview. The checkpoint on Hugging Face is not the one that was being tested: Mistral says it kept scaling reinforcement learning on the preview feedback and saw “massive improvements” in coding and defensive cybersecurity, along with better reasoning efficiency. The model card describes a mixture-of-experts with about one trillion parameters, 49 billion of them active per token (52 billion counting the embedding and output layers), 128 routed experts of which four fire per token plus one shared expert, a one-million-token context window and image input. It was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own data centres in Europe, which is the point Mistral leads with when it sells the model as a foundation for “sovereign AI”.

“Open weight” comes with a condition. The model ships under a new Mistral Ecosystem Open-Weight License, a modified MIT licence that allows commercial and non-commercial use, fine-tuning and redistribution. But a company that runs a “Model as a Service” business, giving third parties API or hosted access with meaningful control over inputs, parameters or training data, must sign a separate commercial licence with Mistral before using the model or its derivatives commercially. Internal use is exempt, and so are products that only embed the model inside a specific feature. In practice that leaves enterprises, governments and researchers free to self-host, while cloud providers and inference start-ups that want to resell Large 4 have to deal with Mistral first.

The coding numbers are where the extra training shows. On Terminal-Bench 4.0, as run by Artificial Analysis, the released model scores 41%, up 13 points from the preview and narrowly ahead of GLM-5.3’s 40%, while using 30% fewer output tokens. On Cognition’s private FrontierCode 2.0, which asks whether a repository maintainer would actually merge an agent’s change, it leads the open-weight field at 19.7%. On the Artificial Analysis Coding Agent Index it scores 48%, which Mistral calls a tie with GLM-5.3, but uses 47% fewer tokens per task. Mistral is also candid about the ceiling: in blind pairwise grading by Surge, Large 4 was preferred over GLM-5.3 and Kimi K3 on 64% of prompts and over DeepSeek V4.1 Flash on 56%, but over Claude Opus 5.5 on only 22%. The model card adds that on the broad Artificial Analysis Intelligence Index v5 it scores 26, the highest of any Western open-weight model, while leading Chinese open models score higher.

The release adds three reasoning settings, low, medium and high, with high as the default and low replacing the old no-reasoning mode, so one deployment can serve both quick chat replies and long agent runs. Outside coding, Mistral reports 73.3% on the FinWorkBench spreadsheet test (second among open models to Kimi K3’s 77.3%), 84.5% on BrowseComp, and resistance to 97.1% of attacks on the B3 agent-security benchmark, up from 93.3% in the preview. Alongside the full checkpoint there is a quantised NVFP4 version for Blackwell GPUs that Mistral says performs the same with a smaller memory footprint. The model also runs on the Mistral API as mistral-large-4 (docs) and on NVIDIA’s build.nvidia.com, and a demo gallery on Hugging Face gives visitors a small free budget during the first week.

Security is the other half of the pitch, and the most hedged. Mistral says Large 4 scores 51% on the Artificial Analysis Cyber Index, which tests finding, reproducing and patching vulnerabilities, and refuses 96.6% of malicious cyber requests across JailbreakBench, StrongREJECT and AgentHarm prompts, the highest rate it measured among open models. Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, Fortinet, Orange Cyberdefense, Rubrik, Zscaler and the EU cybersecurity agency ENISA are quoted as testers. Yet the same post admits that while monitoring preview traffic Mistral saw “a potential for uplift to conduct real attacks,” and says it is changing its guardrails in response. A more permissive variant tuned for penetration testing and red-teaming goes only to vetted design partners such as Orange Cyberdefense, under contract, while everyone else can adapt the public weights with Mistral’s Forge customisation service.

That tension is what the release really tests. Google, by contrast, kept Gemini 4 Carbon away from the public entirely and handed it to vetted cyber defenders first. Mistral has put a trillion-parameter model with a 51% cyber score on a public download page and is relying on trained-in refusals that, once the weights are out, anyone with enough GPUs can fine-tune away. If European banks, telecoms and agencies take it up as the sovereign option it is designed to be, the bet pays off. If the uplift Mistral itself observed turns up in the wild first, Large 4 becomes the case study for every argument against releasing frontier-scale weights.