Mistral Large 4 Goes Open-Weight With 1T Parameters

Mistral has released the weights of Mistral Large 4, a 1-trillion-parameter, natively multimodal model, and says it leads open models on agentic coding while warning that its cyber skills carry “a potential for uplift to conduct real attacks”.

The French company announced the release on its blog on 29 October, after a public preview earlier this month that it followed with three more weeks of reinforcement learning. On X, Mistral said the weights advance “the open frontier for agentic coding and cybersecurity”. The model, nicknamed Le Chonk, is now downloadable from Hugging Face.

According to the Hugging Face model card, Large 4 is a mixture-of-experts model: only part of the network runs for each token. Mistral puts that at 49 billion active parameters per token, or 52 billion counting embedding and output layers. It accepts text and images, handles a context of up to 1 million tokens and, Mistral says, was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own datacentres in Europe.

What Mistral says the model scores

All of the following are Mistral’s own reported results. On Cognition’s FrontierCode 2.0 coding test, Mistral says Large 4 scores 19.7%, the top result among open models. On the Artificial Analysis Coding Agent Index it reports 48%, “effectively tied” with GLM-5.3, which it calls the leading open-weight model, while using 47% fewer tokens per task. Mistral also reports 41% on Terminal-Bench 4.0, 13 points better than the preview with 30% fewer output tokens.

The broader picture is more mixed, and the model card says so. On the Artificial Analysis Intelligence Index v5, Mistral reports a score of 26, which it calls the highest among western open-weight models, while noting that leading Chinese open models score higher. TSN covered the preview’s earlier Artificial Analysis rating when it appeared.

Mistral also cites 73.3% on the Finch finance test, second among open models only to Kimi K3, and 84.5% on BrowseComp, a web research test. It supports 160-plus languages and three reasoning settings, low, medium and high, with high as the default.

The weights come under the Mistral Ecosystem Open-Weight licence, a modified MIT licence. Commercial and non-commercial use is allowed, but businesses selling the model as a hosted service need a separate commercial licence. Through Mistral’s API it costs $0.70 per million input tokens, $0.07 per million cached tokens and $2.10 per million output tokens. Microsoft Foundry is the first cloud provider to carry it, and it is also on NVIDIA NIM and OpenRouter.

Why the cyber warning matters more than the scores

Mistral says Large 4 scores 51% on the Artificial Analysis Cyber Index and refused 96.6% of malicious cyber requests on average across its tests. It then adds an unusually frank caveat: “Though we observed the model refusing to comply with black-box attacks against live targets, we see a potential for uplift to conduct real attacks. We are modifying our guardrails accordingly, and will continue updating them in the months to come.”

That caveat is the most important line in the release. Guardrails on a hosted API can be changed overnight. Guardrails on downloaded weights cannot. Anyone who has the files can fine-tune or strip refusal behaviour, and updates Mistral ships later reach only those who choose to take them. A 96.6% refusal rate also means some malicious requests were answered in testing, before anyone tried to remove the safeguards.

Mistral’s own handling points the same way. It is keeping a specialised cyber variant for vetted enterprise partners such as Orange Cyberdefense, under contractual safeguards and limited to systems they own or are authorised to test. That is close to the route Google took when it gave Gemini 4 Carbon to cyber defenders before developers. The defensive upside is real: Anthropic says an unreleased model found 62 Chrome flaws that have since been patched. But the general Large 4 weights are now public, so the strongest cyber skills in this release sit with whoever downloads them, not only with vetted partners.

For European buyers, the pitch is sovereignty: a frontier-scale model trained in Europe that can run on their own hardware. Mistral has already lined up research partners such as CERN, which signed an intent to build a large physics model with Mistral. What needs watching now is whether independent testers reproduce Mistral’s coding scores, and how Mistral’s promised guardrail updates work for a model it no longer fully controls.