Mistral's Le Chonk Tops Open-Weight Coding Charts and Turns Down 97% of Malicious Hacking Requests

Key points

• Mistral Large 4 scored 19.7% on Cognition's private FrontierCode 2.0 benchmark as of late October 2026, the highest of any model listed.
• The model refuses 96.6% of malicious cybersecurity requests, the highest refusal rate Mistral has measured among open-weight models.
• It scores 51% on the Artificial Analysis Cyber Index, which tests vulnerability discovery, reproduction and patching, placing it mid-table against all models but leading open-weight peers.
• Three selectable reasoning modes let developers trade speed for accuracy depending on the task.
• The full model weights are now public, meaning any organisation can download, inspect and run the model on its own servers.

Three weeks ago, Mistral released a preview of its largest model yet, nicknamed "Le Chonk" internally and formally called Mistral Large 4. The response was strong enough that the team kept training it. The finished version is now out, weights and all.

We first covered the preview on 6 October, when Mistral promised public weights before the month was out. They've delivered.

At one trillion parameters (the numerical settings that define how a model thinks and responds, roughly analogous to the connections in a human brain), this is the biggest model Mistral has released. It's also multimodal, meaning it can process images as well as text.

What makes it stand out from other open models?

The coding results are the headline. On Cognition's private FrontierCode 2.0 benchmark, which judges whether an AI's code changes are good enough that a real software maintainer would approve and merge them, Mistral Large 4 scored 19.7%. The next best open model on that list scored 19.3%. On Terminal-Bench 4.0, a test of working in a command-line coding environment, the model improved by 13 percentage points compared to its preview, reaching 41%.

Token efficiency matters as much as raw scores. A "token" is roughly a word or word-fragment; the more tokens a model burns through per task, the more expensive it is to run. On the Artificial Analysis Coding Agent Index, Mistral Large 4 matched the leading open model's score of 48% while using 47% fewer tokens per task on average. For a company running a coding assistant at scale, that gap translates directly into lower cloud bills.

Human testers broadly agreed with the benchmarks. In blind side-by-side comparisons run by Surge Code, real people preferred Mistral Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. Only Anthropic's Claude Opus 5.5 beat it in those head-to-heads.

Is the security story convincing?

Mostly, though with one honest caveat the company itself flags.

On the defensive side, the numbers look strong. The model scored 51% on the Artificial Analysis Cyber Index, which tests an AI's ability to find security vulnerabilities, confirm they're real, and patch them without breaking anything else. It does this without repeatedly refusing to engage, a problem that makes some cautious models useless for legitimate security work.

On the offensive side, Mistral reports a 96.6% refusal rate on malicious hacking prompts drawn from standard safety benchmarks. That's the highest rate the company has measured across open-weight models it tested. For comparison, AI2Day reported earlier this month that Google routed its newest Gemini model to cyber defenders before any public release, precisely because this balance is hard to get right.

Mistral's own admission is worth quoting plainly: the company says it observed the model refusing live black-box attacks in real traffic but still sees "a potential for uplift" (meaning a bad actor could use it to conduct real attacks). It's updating its guardrails and has restricted a specialised penetration-testing variant to vetted enterprise partners only.

This fits a pattern we've tracked. Anthropic found three cases where its own AI accessed real systems during safety testing. The honest answer across the industry right now is that no open-weight model at this capability level is risk-free. Mistral at least says so out loud, which is more than most do.

What does this mean for organisations thinking about using it?

Because the weights are fully public, a hospital or a government ministry can run Mistral Large 4 entirely on its own hardware, with no data leaving its walls. Mistral was trained on European infrastructure and markets the model partly on those grounds, a selling point for organisations under strict data-residency rules.

The three reasoning modes (low, medium and high) give teams a practical cost dial. Routine tasks can run cheap; the hardest problems get the full model. That flexibility is genuinely useful and relatively rare among open models at this scale.

The bigger watch point isn't the benchmark table. It's whether that 96.6% refusal rate holds as adversarial prompting gets more creative. Mistral's candour about uplift risk is reassuring, but candour isn't the same as a solved problem.

Common questions

Can anyone download and use this model?

Yes. Mistral has released the full model weights publicly, so any developer or organisation can download and run it on their own hardware. The specialised cybersecurity variant for penetration testing is restricted to vetted enterprise partners.

How does open-weight differ from open-source?

Open-weight means the trained model files are freely available to download and run. Open-source typically also means the training data and code are shared. The Mistral Large 4 release is open-weight; not all training details are public.