Mistral Opens the Weights on Its Trillion-Parameter Large 4, Pitching It to Cyber Defenders

Mistral Large 4 open weights land on Hugging Face

Mistral AI released the open weights of Mistral Large 4, nicknamed Le Chonk, on October 29, 2026, after an API-only public preview on October 6. The Mistral Large 4 open weights are downloadable from Hugging Face for self-hosting, which is the whole point of the release. Anyone with enough hardware can run the model themselves, fine-tune it, and build on it without routing every request through Mistral's servers. That is a real difference in a market where the biggest models usually stay behind an API, and it is why this release matters even if you never touch a trillion-parameter model yourself.

The model is a mixture-of-experts design with about 1.05 trillion total parameters and roughly 52 billion active per token. It takes text and image input, produces text output, and carries a 1 million token context window. In plain terms, it can chew through very large documents and codebases in a single pass, and it only activates a fraction of its weights for each token, which keeps the compute bill from scaling with the headline parameter count. Mistral trained it from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, a detail the company is keen to advertise.

The license is open, with one commercial catch

The weights ship under a modified MIT license, the Mistral Ecosystem Open-Weight License. You can use, fine-tune, and commercialize the model freely. The one catch is for companies that want to resell it as a model-as-a-service offering, which needs a separate commercial license. For everyone else, including a security team that wants to run it inside its own network, the terms are permissive. Distribution is broad from day one: Mistral API and Mistral Studio, OpenRouter, Nvidia NIM containers, and Microsoft Foundry as the first cloud provider. API pricing starts at 0.70 dollars per million input tokens and 2.10 dollars per million output tokens.

That distribution matters because open weights only help if you can actually run the thing. A trillion-parameter model is not a weekend project on a laptop. The quantized checkpoints and NIM containers are what turn a research artifact into something a well-resourced team can deploy, and Mistral clearly wants this model in production quickly, not just on leaderboards.

The security pitch is the headline

Mistral is pitching Large 4 first and foremost to cyber defenders. Its own numbers claim 51 percent on the Artificial Analysis Cyber Index and an average 96.6 percent refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT, and AgentHarm. The idea is a model that is strong at finding vulnerabilities and writing fixes, but refuses to help attackers do the same. A more aggressive penetration-testing version stays restricted to vetted partners, which tells you Mistral knows the dual-use problem here is real.

That framing is worth comparing with Google's approach to its own security model. Google rolled out Gemini 4 Carbon slowly, to trusted defenders first, for exactly the same reason: a model good at finding flaws is useful to both sides. Mistral is making a different bet. It argues that defenders gain more from running the model themselves, under their own policies, than from waiting behind a provider's filters. If your security work involves sensitive code or classified networks, you cannot send it to an outside API at all. Self-hosting is not a preference in that world. It is a requirement.

The benchmark claims need independent testing

Mistral's other benchmark claims are strong, and they should be treated as vendor claims until independent labs confirm them. The company reports 48 percent on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47 percent fewer tokens per task, 19.7 percent on Cognition's FrontierCode 2.0 (the top open-weight score by its account), and 41 percent on Terminal-Bench 4.0, up 13 points from the preview. Those are impressive numbers, and the token-efficiency claim is arguably the most important of them: a model that does the same work with half the tokens is cheaper to run at scale, which is what turns a demo into a deployment.

But benchmarks are a controlled environment, and coding benchmarks in particular have a habit of flattering the model that was tuned closest to them. The honest test is a team's own codebase, its own tickets, and its own definition of done. That is also where the open-weights model has a structural advantage over the closed kind: you can run that test yourself, on your own data, without asking anyone's permission.

What this means for the open model race

Large 4 arrives into an open ecosystem that is settling into distinct roles. The open source AI stack already has vLLM, Ollama, and llama.cpp each owning a job, and a model like this slots into the serving layer that stack provides. Where OpenAI's GPT-6.1 Terra pitches cheap, high-volume work, Mistral pitches sovereignty and control: European training infrastructure, weights you can download, and a license that lets you build a business on top.

One caveat from the release itself is worth noting. At least one digest reported that the main Hugging Face checkpoint page still showed an upcoming release with an October 31 ETA, while a quantized checkpoint was already available for self-hosting. If you go looking for the weights, check which checkpoint you are actually downloading. The direction of travel, though, is clear. The frontier is no longer only accessible through an API key. For teams that need control over their data, their policies, and their costs, that is a genuine shift, and Le Chonk is the biggest open-weight statement of it yet.