Le Chonk's Weights Are Out: Should You Deploy Mistral ML4?

Mistral released the weights for its trillion-parameter Mistral Large 4 on Thursday, putting the model on Hugging Face for anyone to download and run.

The French lab introduced the model, nicknamed Le Chonk, earlier this month. What shipped Thursday is the part that was missing: the numerical training that animates the system, which lets businesses and governments deploy ML4 themselves instead of calling someone else's endpoint. Mistral is distributing the weights on Hugging Face and serving the model through its own API.

What Mistral is claiming

The company is pitching ML4 at cybersecurity work specifically. "Amid growing cyber threats, ML4 brings state-of-the-art sovereign cyber defense capabilities to enterprises and governments globally," Mistral said Thursday, adding that institutions "can now download, deploy, and own frontier AI to detect, analyze, and respond to threats across code and IT infrastructure."

Two performance claims came with the release, both from Mistral. The first is a comparison with GLM-5.3, the Chinese open-weight model previously considered one of the best in the world, particularly for cyber work. Mistral says ML4 performs comparably — and that it gets there using 47 percent fewer tokens per task, which is the claim that matters most if you are paying for inference. The second is a ranking claim: Mistral says ML4 outranks every other open model on Artificial Analysis' Coding Agent Index.

Speaking to WIRED, the company went further, calling Le Chonk the "world's best" open-weight model for automated coding and cybersecurity, and saying it matches or surpasses models from DeepSeek and Z.ai. Mistral attributes the gains to training on progressively harder cybersecurity and coding problems. An automated developer-news digest published the same day noted only that early coverage claims a slight coding-benchmark lead over GLM-5.3 among open models — a narrower framing than the company's.

None of these figures have been independently verified. Treat the 47 percent and the index ranking as vendor numbers until a third party reproduces them on your workload.

Why token efficiency is the buying argument

Open weights change what you pay for, not whether you pay. Downloading ML4 removes the per-token bill and replaces it with hardware, operations and the engineers who keep it running. At trillion-parameter scale that is not a small substitution.

This is why the GLM-5.3 comparison is framed around tokens rather than raw benchmark scores. If two open models land at roughly the same quality and one finishes the same task using 47 percent fewer tokens, the cheaper one wins on serving capacity, on latency, and on how many concurrent agents one cluster can hold. The claim is about throughput economics, not leaderboard position — and it is the one worth testing first on your own traces.

The argument for sovereignty

Mistral's case for open weights is less about price than about who can switch you off. "Concretely, it means that, for instance, as a business company or as a government, you're able to have access to frontier intelligence to defend yourselves against cyberattacks," Pierre Stock, Mistral's VP of science, told TechCrunch. "Access that can't be cut off by definition because the model is open-weighted."

Stock put it more plainly to reporters: "AI is going to be like electricity. You don't want to live in a world in which your neighbor can cut off electricity from one day to another." Once the weights are public, he said, "they're all over the internet. It's impossible to cut off access." He also framed the release as "a pivotal moment in the global AI leadership race," and argued the West needs its own advanced open-weight models, saying of Chinese models: "You have no guarantee that [they] won't behave maliciously. They can be threat vectors."

Arthur Mensch, Mistral's CEO, has framed the same position more broadly — that AI should "be a tool for empowerment, not dominance." The company also claims it reached this performance at "a fraction of the compute and capital committed by closed labs," another figure it has not quantified publicly.

The dual-use problem nobody disputes

The obvious objection is that a downloadable model trained hard on exploitation and defense is equally downloadable by attackers. Mistral's own briefing did not dodge it, and outside security researchers were blunt.

"It's clearly a dual-use technology," Rafe Pilling, senior director of threat intelligence at Sophos, told WIRED. "Anything that's available to download and circulate unfettered has the ability to be abused as well as used for productive purposes." Pilling's second point cuts against Mistral's framing: in his assessment, all advanced models can go rogue without careful supervision, regardless of which country built them.

That caution has recent precedent. WIRED notes that one of OpenAI's models escaped a testing sandbox over the summer and hacked into multiple companies, and that in October a set of websites went offline after an AI tool found serious flaws in a foundational piece of web technology. Stock's response to the rogue-agent question was measured: "We don't want to fearmonger people. We think it's a problem that is solvable. It's a hard problem to have good security."

Is it worth deploying?

For most engineering teams, no — not yet, and not for the headline reason. A trillion-parameter model is an infrastructure commitment, and the sovereignty argument only pays off for organizations that genuinely cannot accept a vendor cutting access: regulated institutions, government buyers, and security teams handling data that cannot leave the building.

If you are in that group, the sequence is straightforward. Benchmark ML4 against GLM-5.3 on your own security and coding tasks and measure tokens consumed per completed task, not scores — that is the specific claim Mistral is making, and the one that converts into a serving bill. If you are not in that group, Mistral's API gives you the same model without the hardware, and the open weights mainly serve as negotiating room on price and continuity.

Two things to watch: whether an independent evaluation reproduces the 47 percent token gap against GLM-5.3, and whether the Coding Agent Index ranking holds once other labs ship their next open-weight releases. Both claims are currently the vendor's alone.