Mistral has released the full weights for a trillion-parameter model it’s nicknamed Le Chonk, and the company is pitching it specifically as a cyber defense tool governments and enterprises can run entirely on their own servers.
TL;DR: Mistral Large 4, which the French AI company first previewed earlier in October, is now downloadable on Hugging Face as an open-weight model, meaning anyone with the hardware can deploy it themselves instead of calling an API. Mistral is framing the release around cybersecurity specifically, arguing that defenders shouldn’t have to depend on outside companies for frontier AI capability they might need in the middle of an active attack. The company claims performance on par with GLM-5.3, a leading Chinese open-weight model, while using fewer tokens per task.

What “open weights” actually buys you here
There’s a real difference between a model you access through an API and one you can download and run yourself. With weights in hand, an organization can deploy Mistral Large 4 entirely inside its own infrastructure, with no external dependency that could be throttled, priced differently, or simply shut off. For a cybersecurity team, that’s not a minor convenience. Pierre Stock, Mistral’s Vice President of Science, made the case directly to TechCrunch: open weights mean access to a frontier model for cyber defense that cannot be cut off, and organizations can run it entirely on-premises inside their own environments.
“We don’t want to fearmonger people,” Stock said, addressing concerns about rogue AI agents, while still acknowledging the risks are real.
CEO Arthur Mensch framed the broader philosophy behind the release in similar terms, arguing AI should function as a tool for empowerment rather than a mechanism of dominance by whichever lab holds the most capable closed model at any given moment. He said Mistral’s approach reaches state-of-the-art performance using a fraction of the compute and capital committed by closed labs, a claim that’s become something of a standing argument among open-weight developers positioning themselves against OpenAI and Anthropic.
The performance claims, with the caveat attached
Mistral says Large 4 performs comparably to GLM-5.3, a Chinese open-weight model that outlets including Fortune have already flagged as one of the strongest open options available, particularly for cyber-related work. Mistral’s specific claim is that Large 4 uses 47% fewer tokens per task than GLM-5.3, which the company presents as a meaningful efficiency and cost advantage rather than just a capability claim. Mistral also says the model leads all other open models on Artificial Analysis’ Coding Agent Index, a benchmark focused specifically on agentic coding performance.
None of those figures are independently verified yet, and Mistral hasn’t published detailed benchmark scores or licensing terms alongside the release. That’s worth sitting with for a moment: a company asking enterprises and governments to trust a model for cyber defense work is, for now, asking people to take its own efficiency and performance comparisons largely on faith until outside researchers get a chance to run their own tests.
The sovereignty argument, and why it lands differently in security
Mistral’s pitch about sovereign AI isn’t new, the company has leaned on European data independence arguments since its founding. What’s different here is applying that logic specifically to cybersecurity, where the argument genuinely carries more weight than it does for, say, writing marketing copy. A government agency defending critical infrastructure has a much stronger reason to want a model it fully controls and can audit, rather than one living behind someone else’s API that could, in theory, be modified, monitored, or revoked at a moment that matters.
Mistral’s own statement leaned into this directly: “Amid growing cyber threats, ML4 brings state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally.” It’s a pitch tailored almost perfectly to European regulators and governments who have spent the past few years worried about depending on American AI infrastructure for anything security-sensitive.
The double-edged nature of handing out a trillion-parameter model
Stock’s comments about malicious actors weren’t throwaway lines. Any model capable enough to help a defensive security team is, almost by definition, capable enough to help an attacker who downloads the same weights. Mistral’s position is essentially that getting strong defensive AI tools into the hands of legitimate organizations quickly matters more than the risk of bad actors also getting access, on the logic that sophisticated attackers will eventually get access to comparably powerful tools regardless of what any single company decides to release. Whether that tradeoff holds up is going to depend heavily on how Large 4 actually gets used in the wild over the next few months, not on anything in Mistral’s launch messaging. The stakes are real: the same week this released, researchers were still untangling the fallout from a fake Claude Code extension that planted SSH backdoors on thousands of developer machines, a reminder that AI coding tools are already a live target regardless of which lab built them.
What enterprises still have to figure out themselves
Downloading a trillion-parameter model is a very different undertaking than calling an API, and that gap is worth being honest about. Running Large 4 on-premises means an organization needs the actual hardware to host it, the engineering talent to deploy and maintain it securely, and a plan for keeping it updated as Mistral ships improvements. None of that is trivial, and it means the sovereignty pitch mostly lands with larger enterprises and governments that already have serious infrastructure budgets, rather than smaller security teams who might benefit just as much from the capability but can’t realistically stand up the hardware to run it themselves.
Where this fits in the open-weight race
The open-weight model space has gotten genuinely competitive this year, with Chinese labs producing models like GLM-5.3 that Western developers increasingly have to benchmark themselves against rather than dismiss. Mistral positioning Large 4 as both commercially available through its API and freely downloadable is a bet that enterprises want the choice of deployment model more than they want to be locked into one or the other, and that a cybersecurity angle gives it a clearer story than just competing on general-purpose benchmarks. It’s a different approach from GitHub’s recent Copilot Fusion model router, which leans into choosing between closed models rather than betting on one open one.
The bottom line
Handing security teams a trillion-parameter model they can run entirely on their own hardware is a genuinely useful idea, and Mistral’s sovereignty argument is a real one, not just marketing. The catch is that every performance number in this release is Mistral’s own, and the same openness that lets a defender deploy it without restriction lets anyone else do the same. Whether Le Chonk turns out to be mostly a defensive tool or mostly a wash depends on who downloads it next.



