Mistral Large 4 open weights: what the cyber pitch offers and what's still unproven

Verdict box

What it is: Mistral Large 4 (nicknamed "Le Chonk"), a 1-trillion-parameter model whose weights you can now download.
Best for: security teams and agencies that want to run a frontier-class model on their own hardware.
Skip if: you only want a chatbot, or you don't have GPUs to spare.
Scores: I haven't used it, so I'm not giving it features, ease of use, value or performance ratings. Everything below comes from Mistral and other people's reporting, and I say which is which.

Let me be upfront. This is a review-category piece, but I haven't run this model. A 1-trillion-parameter model isn't something you spin up on a laptop, and I'd rather tell you that than fake a test. What I can do is read the claims, check them against other outlets, and tell you where they're thin.

What it is

On October 29, TechCrunch reported that French AI company Mistral had released the weights for Mistral Large 4 (ML4), letting businesses and governments download and deploy the model themselves. The model was first introduced earlier in October. CNBC covered that October 6 unveiling, where Mistral said it was particularly strong at cyber, coding, manufacturing, finance and multimodal tasks.

"Weights" are the trained numbers that make the model work. Having them means you can host the model yourself instead of calling someone else's API. That's the whole pitch.

According to The Next Web, the model reads text and images and produces text and code. Mistral says it was trained from scratch in Europe. Mistral has also written up the release on its own site.

One small mismatch between sources: The Next Web says 52 billion parameters are active per task, while VKTR's preview coverage put the figure at 49 billion. The Hugging Face repository name Mistral links to includes "A52B", which fits the higher number.

The cyber argument

Mistral's case is about control. Pierre Stock, its VP of Science, told TechCrunch that organizations can have frontier intelligence for defense, "access that can't be cut off by definition because the model is open-weighted." He also said companies can deploy it on-prem in their own environment.

That's a real point for some buyers. If your security tooling depends on a closed model, the provider can change terms, pricing or policies. Cloudflare's chief strategy officer Stephanie Cohen makes a similar case in Mistral's announcement, saying defenders should be able to audit the architecture and run it on their own infrastructure.

Keep in mind that quote sits on Mistral's own page, alongside praise from other partners. It's a marketing page, not an independent review.

What's been measured (not by me)

Here are the numbers on the table. Most come from Mistral's own charts, so treat them as vendor claims until more outsiders publish.

Read that cyber row carefully. TechCrunch relays Mistral's claim that ML4 is comparable to China's GLM-5.3 on cyber work. But on the Cyber Index chart on Mistral's own page, the 51% sits below a few other open models. So "best open model for cyber defense" depends heavily on which benchmark you pick. The coding story looks stronger than the cyber one.

The efficiency claim is the more interesting one. TechCrunch says the 47% token saving makes ML4 significantly cheaper to run than GLM-5.3. That matters if you run agents for hours.

Reasoning modes and hardware

ML4 now has three reasoning modes: low, medium and high, with high as the default. Mistral recommends low for simple tasks, medium for everyday work and high for the hardest jobs. The Next Web says early testers asked for more control over how long the model thinks.

On hardware, The Next Web says Mistral is also releasing a quantised version that runs on four Nvidia B200 or B300 GPUs. That's still a serious data-center bill. Smaller distilled versions, including one meant to run on a laptop, are promised "in the coming months", but Mistral wouldn't commit to dates.

Pricing and license

Sources disagree here. Mistral's announcement says that starting today ML4 costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The Next Web, published the same day, says API pricing hasn't been published yet. VKTR listed preview pricing at $1.36 input and $4.18 output per million tokens. I'd trust Mistral's page for its own price list, but check the pricing page before you budget.

The license is a modified MIT license, according to Mistral. It's free to use, modify, fine-tune and commercialize, with one condition: businesses that run a model-as-a-service offering need a separate commercial license from Mistral.

Where can you get it? Mistral lists its own Studio, its API, OpenRouter, Hugging Face (for the weights), Nvidia-hosted endpoints and Microsoft Foundry.

What other reviewers and users say

Not much yet, and I won't invent a consensus. The weights went live today. I found news coverage from TechCrunch, The Next Web and others, but no hands-on video reviews or community verdicts I could responsibly cite. Mistral itself asks the community to run its own evaluations and share findings. That's the right call. Wait for a few of those before you trust the leaderboard.

Mistral does cite a blind human-preference test from Surge Code, where graders preferred ML4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash. In that test it ranked behind only Claude Opus 5.5 among the models evaluated. Again, that's Mistral reporting it.

The safety question

An open-weight cyber model cuts both ways, and Mistral admits it. In its announcement, the company says it watched cyber-related use in live traffic after the preview. The model refused black-box attacks on live targets, but Mistral still saw potential for uplift in real attacks and is changing its guardrails.

Here's the catch: once weights are public, you can't recall them. Stock argued that open release helps defenders get capability fast, given that some malicious actors may use AI to attack. The Next Web notes that rivals argue open models carry more risk. Both views are on the record, and I can't settle it for you. Mistral is also giving vetted design partners such as Orange Cyberdefense preview access to a specialized cyber variant, with testing limited to authorized systems.

Pros and cons

Pros

• You can host it yourself, so no one can cut off access.

• Strong coding results among open models, with notably fewer tokens per task.

• Permissive license for most businesses.

• Three reasoning modes let you trade cost against accuracy.

Cons

• Running it yourself needs serious GPU hardware.

• Cyber benchmark leadership isn't clear-cut, even on Mistral's own charts.

• Most numbers are self-reported and brand new.

• Pricing reports conflict, and the laptop-sized versions have no date.

• Public weights can't be recalled if misuse grows.

Who should look at it

If you run a security team, a public agency or a regulated company that cares about data residency and vendor lock-in, ML4 is worth a trial on your own workloads. That's the audience Mistral is courting, and its partner list leans that way. If you just want the best chatbot, nothing here is built for you.

My verdict, for now: promising, mostly unproven. The efficiency and coding numbers are the strongest part of the case. The cyber leadership claim needs independent testing. I'll update when outside evaluations land.