In cyber defense, owning the model
can beat leaderboard rank

Mistral just gave away a trillion-parameter model. On October 29, 2026 ⁰⁵, it released the weights for Mistral Large 4 on Hugging Face. Mistral says the model holds 1.05 trillion ⁰⁷ parameters, yet only 52 billion ⁰⁹ fire on each token. That is about 5% ²⁵ of the model doing the work at any moment.

It is probably not the smartest model you can rent. Google unveiled Gemini 4 Carbon one day earlier and says Gemini 4 Carbon (High) leads the Artificial Analysis Intelligence Index v5.0 at 59 ⁰¹. Mistral reports a 51% ⁰² score for Large 4 on a different test, the Artificial Analysis Cyber Index. Those numbers come from two separate indexes, so subtracting one from the other tells you nothing.

We will start with an admission. Nobody has published a same-test score showing which of the two models finds more bugs. For a security team, that missing comparison matters less than a simpler question. Where does your data go when the model reads it?

A rented model reads your unpatched bugs on someone else's servers. A downloaded model reads them inside your own walls. For defenders, that difference can be worth more than a few leaderboard points.

Control-Adjusted Performance Beats Raw Rank

A benchmark score tells you how well a model handles the work it is allowed to see. Control-adjusted performance asks a second question: how much of your real work is it allowed to see? Multiply the two and you get the number a security buyer should actually care about.

Try it with two made-up models. Model A solves 60% ²⁶ of tasks, but your lawyers let it touch only half your code, so it handles 30% ¹⁰ of the real workload. Model B solves 50% ²⁷ of tasks and can see everything. It handles 50%.

The day's facts sort into three parts of control. The first is where the data lives. Mistral pitched the release at enterprises and public-sector bodies that want to deploy and own frontier AI for threat detection. A self-hosted copy runs in your own data center, so source code and incident logs never cross to a vendor's API.

The second is what you can change. With the weights, a team can fine-tune on its own incident history. Fine-tuning means extra training on your data, so the model learns your logging format and your codebase. A closed model leaves you with prompts plus whatever tuning the vendor sells.

The third is what you can freeze. You can pin one version and keep it. When a hosted model changes behind an API, your team may watch behavior shift with no old copy to compare against. A pinned copy lets you approve upgrades through your own change process.

All three come with a bill. NVIDIA's model card lists about 49 billion ²¹ active parameters per token, or 52 billion ⁰⁹ counting embeddings and output layers. Sparse or not, a trillion parameters still have to sit in memory somewhere. You now own the servers, and the people who patch them.

What Le Chonk Proves So Far

Mistral nicknamed the model "Le Chonk," and its track record is short but moving fast. The public preview opened on October 6, 2026 ²⁰. By October 29, Mistral says its Terminal-Bench 4.0 score rose from 28% ⁰³ to 41% after more reinforcement learning. That is 13 ¹⁰ points in 23 days, and by Mistral's count output tokens on Terminal-Bench 4.0 fell 30% over the same stretch.

On cyber, Mistral's own reporting looks strong. Mistral reports that matches MiMo-V2.6-Pro and beats DeepSeek V4.1 Flash at 76% ¹³. NVIDIA's model card reports 48% ⁰⁴ on the Artificial Analysis Coding Agent Index.

Vendor numbers deserve a discount, so look at an outside test. Averlon, a security company, ran Large 4 through 16 ¹⁸ security benchmarks. It landed in the upper half on 12 ¹⁴, including all eight where the evidence sat in front of the model or could be searched. It tied for last on misconfiguration triage. At list price, Averlon placed it in the most expensive third of runs on most benchmarks.

That pattern is the real lesson. Averlon found Large 4 strong when handed the evidence and weaker when it had to recall specifics from memory. Security teams own exactly that evidence. Mistral's docs list a 1-million-token context window, enough to fit a large codebase or a long incident timeline in one prompt.

Cash flow decides who should self-host. Mistral's API lists promotional public-preview prices of $0.70 ⁰⁶ per million input tokens and $2.10 per million output tokens. Self-hosting trades that metered bill for fixed costs: GPUs, and the engineers who run them. A government agency running millions of sensitive queries may justify its own cluster. A small company with bursty usage will likely spend less on the API.

The attacker gets the same download. Fortune reports that some experts expect capable open models to enable cheaper, scalable agentic attacks. Anthropic's chief executive has warned that open models could contribute to widespread cyberattacks, per Fortune. Mistral spent its preview red-teaming with vetted partners and state authorities, who got reduced moderation and expanded cyber capabilities.

We think the control case wins for regulated defenders holding data that cannot leave the building. For most small teams, the API is the better first step. Whether Averlon's rankings hold once a team fine-tunes the model on its own incidents is an open question. Nobody has published that test yet.

The evidence supports one decision today: start measuring. Three things would change the call. A same-index comparison against Carbon, a license term that blocks your use, or a poisoned fine-tune circulating under Mistral's name. NVIDIA's model card tells anyone who wants to deploy and customize the model in their own environment to contact Mistral. Read the Mistral Ecosystem Open Weight License before you build on it.

Where a trillion open parameters help and hurt

Owned Security Models Go Mainstream by 2031

The release cadence is tight. Google shipped Carbon on October 28, and Mistral shipped weights on October 29. Mistral says its gains "continue to accelerate, with no signs of saturation." That is a vendor claim until outside tests repeat it.

Cost per task is easier to project. Say a security agent run burns 1 million ²⁹ output tokens on Mistral's API, which costs $2.10 ⁰⁶. Mistral's 30% ¹⁰ token cut brings the same job down to 700,000 tokens ³⁰, or $1.47. If one more cut that size landed each year through 2031, the run would cost about 25 ³¹ cents ($1.47 × 0.7 × 0.7 × 0.7 × 0.7 × 0.7).

Once model cost falls toward cents, the scarce asset moves. It becomes your answer key: labeled incidents, plus the harness that scores models against them. Building that harness is cheap, and it pays off with any model, open or closed. Skip it and you stay locked to whichever vendor you picked first.

The bet is lopsided. The downside is a few weekends of engineering time. The upside is the freedom to swap in whatever model wins next quarter, including a self-hosted one when your data rules demand it.

Mistral cannot outspend Google on leaderboard rank, so it competes on ownership. Our read is that an API-first vendor will struggle to copy this counterpositioning, because giving away its best weights would undercut its own metered business. Security procurement may start scoring sovereignty as a line item by 2031, the way it scores uptime today. The data on whether buyers will pay that premium is still thin.

Score Large 4 on Your Own Fixed Bugs

Step one: pick ten vulnerabilities your team fixed in the last year. You already know the right patch for each one. Together they form your answer key.

Step two: check out the commit from just before each fix into a sandbox. A sandbox is an isolated copy with no production credentials and no network path to live systems. The model gets read-only access to the repository and nothing else.

Step three: run Large 4. If your code is allowed to leave, call it as mistral-large-4-0 on Mistral's API. If it is not, run the same test this weekend on a public C or C++ project with known, already-patched bugs. Self-host later, once the numbers justify the hardware.

Step four: log four numbers per bug. Record found or missed, whether the patch passes your tests, output tokens used, and minutes of human review. Multiply output tokens by $2.10 ⁰⁶ per million for the raw model cost, then divide total cost by the patches that passed.

Step five: run one closed model through the identical harness. Use the same prompts and the same scoring. Expect breakage along the way. A run that invents a file path or a patch that fails a test is useful data, so write it down.

Step six: rerun the set every month. Pinned weights turn this into a regression test, a check that the model still does what it did last month. When the next model ships, you score it in an afternoon on your own bugs.