Back in June, the internet invented a ridiculous fake Mistral model called “Le Chaton Fat,” complete with absurd specs, fake benchmark charts, and claims that it could crank out “1,000 meows per second.”
We covered the joke at The Neuron. Now Mistral has done something funnier: it built a real monster model and nicknamed it “Le Chonk.”
The joke lasts about five seconds. The business question underneath it is much more interesting.
On October 6, Mistral launched Mistral Large 4, or ML4, in public API preview. You can use the model through Mistral today. By the end of October, the company says it plans to release the model’s weights, potentially giving customers far more control over how ML4 is deployed—depending on the final license and infrastructure requirements.
That makes ML4 an unusually useful experiment in one of AI’s biggest corporate buzzwords: sovereignty.
Because Mistral is effectively asking companies to value something beyond intelligence.
It is asking them what control is worth.
- The benchmarks are good. That’s almost the boring part.
- Mistral wants to change who holds the keys
- “Sovereign AI” has layers
- And right now, that premium is visible in the numbers
- Cybersecurity is where this gets uncomfortable
- Before calling this “AI sovereignty,” ask five questions
- The real test starts when the weights arrive
The benchmarks are good. That’s almost the boring part.
ML4 is a roughly one-trillion-parameter mixture-of-experts model, meaning it contains a huge amount of capacity but activates only a fraction of it for any given task.
Mistral is pitching it aggressively across coding, cybersecurity, finance, legal work, and visual reasoning. Some of those claims still come from Mistral’s own evaluations, so they deserve the usual benchmark-sized grain of salt.
Independent testing, however, suggests there is real substance here.
Artificial Analysis gave the preview a score of 38 on its Intelligence Index. ML4 scored 50 on its Cyber Index, tying GLM-5.3-Flash and trailing MiMo-V2.6-Pro at 56. Its standout result was 82% on CyberGym-E2E-AA, a benchmark focused on reproducing and patching software vulnerabilities. Artificial Analysis’ independent evaluation provides the full comparison.
That is enough to make ML4 a serious model.
It does not establish that Mistral has built the best AI model overall, or that Europe suddenly sits at the top of the global leaderboard. Benchmark results vary dramatically depending on the task, model configuration, agent harness, and competitors being measured.
More importantly, squeezing another few points out of a benchmark is not ML4’s most interesting proposition.
The bigger product is permission.
Mistral wants to change who holds the keys
Many leading proprietary AI services operate like rented infrastructure.
You send a request to somebody else’s servers. They run the model. They decide which versions remain available, what the acceptable-use rules are, which requests get blocked, what prices change, and—in extreme cases—whether you continue getting access at all.
Downloadable weights can change that relationship.
If Mistral’s final release gives customers sufficient rights and the model proves practical to operate, organizations could gain much more control over where ML4 runs, how it is customized, which systems it touches, and how dependent they remain on Mistral’s own API.
That distinction becomes increasingly important as companies move AI deeper into workflows involving proprietary data, critical infrastructure, software development, finance, and security.
Mistral’s pitch is that customers should be able to control more of that stack themselves.
But there is an important terminology trap here: open-weight does not automatically mean open source.
The Open Source Initiative’s definition of open-source AI goes further, requiring the information, code, and model parameters needed to study and modify the system and build a substantially equivalent one. Downloadable weights alone do not meet that standard.
And in ML4’s case, the public weights have not arrived yet. Neither have the final license terms that will tell customers exactly what they are permitted to do with them.
So for now, Mistral is selling the possibility of greater control.
The interesting part begins when buyers try to exercise it.
“Sovereign AI” has layers
Mistral says ML4 was trained in its own European data centers using NVIDIA hardware. That gives the company a strong European-infrastructure story, especially for organizations concerned about jurisdiction, data location, or reliance on foreign AI providers.
But sovereignty is layered.
A company can gain independence from one API provider while remaining dependent on NVIDIA hardware. It can host a model locally while still relying on outside software, engineers, updates, security tooling, and infrastructure suppliers.
And having permission to download a model does not mean you can practically run it.
Think of an API as renting a furnished office. Someone else handles the building, power, maintenance, and security. Self-hosting gives you the building keys—but congratulations, the boiler is now your problem too.
That is why ML4’s eventual license and deployment requirements matter almost as much as the model itself.
If only a handful of large enterprises, governments, and infrastructure companies can afford to operate it efficiently, open weights could create considerably more autonomy without creating much broader access.
Control could become a premium feature.
And right now, that premium is visible in the numbers
Artificial Analysis found ML4’s standard cost per Intelligence Index task was about $1.13.
Its launch promotion temporarily cuts that to about $0.57, but comparable results in the evaluator’s testing cost roughly $0.25 for GLM-5.3-Flash and $0.27 for DeepSeek V4.1 Flash. Mistral’s listed API pricing is $1.36 per million input tokens and $4.18 per million output tokens before its temporary launch discount.
That does not automatically make ML4 a bad deal.
An enterprise might happily pay more for greater control over deployment, data, policy, or continuity. The cheapest benchmark run and the best enterprise purchase are not necessarily the same thing.
But it makes one part of the sovereignty pitch impossible to ignore: independence has an economics problem.
And API prices are the easy costs to measure.
We still do not have a clear public picture of the hardware, energy, engineering time, inference optimization, security, and maintenance required to run ML4 efficiently outside Mistral’s infrastructure.
That means the most important number in this launch may be one Mistral cannot put on a benchmark chart yet: the total cost of actually taking control.
Cybersecurity is where this gets uncomfortable
Mistral makes its strongest philosophical argument around cybersecurity.
Security researchers regularly work with malware, exploits, vulnerability reproduction, penetration testing, and other material that can look suspicious to an AI safety system even when the work is legitimate.
Mistral argues that defenders cannot always depend on centrally controlled models whose moderation systems may reject those requests. During ML4’s preview, the company says cybersecurity leaders, vetted partners, and state authorities are testing the model with reduced moderation and expanded cyber capabilities.
Independent testing suggests ML4 really is capable in this area. That 82% CyberGym-E2E-AA result is noteworthy.
But it is an 82% benchmark result. It does not mean ML4 prevents 82% of cyberattacks, makes organizations 82% safer, or produces a net security benefit once comparable capabilities become more widely accessible.
The same flexibility that makes a powerful model attractive to defenders creates a harder governance problem when organizations control more of the system themselves.
Who decides who gets access? Who monitors misuse? Who patches vulnerabilities? Who responds when something goes wrong?
For a self-hosted system, more of that operational burden—access controls, monitoring, updates, incident response, and governance—moves to the organization running it.
That matters even more in Europe now that the European Commission’s enforcement powers for general-purpose AI obligations have been active since August 2, 2026.
Mistral is a signatory to the EU’s General-Purpose AI Code of Practice, but signing the code is a framework for demonstrating compliance—not proof that every future model or deployment is automatically safe or compliant. The European Commission’s guidance on general-purpose AI also makes clear that obligations depend on the actor’s role and how a model is provided or modified.
Open weights redistribute power.
They also redistribute work.
Before calling this “AI sovereignty,” ask five questions
For companies watching ML4, the useful questions are surprisingly practical:
- What does the final license actually permit? Downloading a model is much less useful if important commercial, modification, or redistribution rights remain restricted.
- What does production deployment require? Hardware, memory, throughput, latency, and serving requirements determine whether self-hosting is realistic outside very large organizations.
- What is the total cost of ownership? API prices are easy to compare. GPUs, engineers, electricity, monitoring, updates, and security are not.
- Which kind of control do you actually need? Data residency, provider independence, customization, moderation discretion, jurisdiction, and supply-chain independence solve different problems.
- Who owns the operational burden? Running more of the stack yourself also means deciding how access, abuse monitoring, updates, incident response, and model governance work.
Those questions matter more than whether ML4 moves three places up or down a leaderboard next week.
And they apply far beyond Mistral.
Earlier this year, we looked at how open models are changing the economics of “renting” versus owning AI. We also examined why downloadable models create a very different security problem from covertly extracting capabilities through somebody else’s API.
ML4 pushes both arguments deeper into the enterprise market.
The real test starts when the weights arrive
Mistral has already demonstrated something meaningful: a European AI lab can produce a model that independent evaluators consider competitive in important areas, including cybersecurity.
That gives enterprises another credible option in a frontier-model landscape where many of the strongest performers have come from American and Chinese labs.
But ML4’s bigger promise still has to survive contact with reality.
The real test comes when a customer can take the released model, inspect the license, calculate the infrastructure bill, run it on systems it controls, secure the deployment, and decide whether the extra responsibility is worth the independence.
If organizations can do that at a price they can justify, Mistral will have created more than another strong model. It will have changed the bargaining position between AI providers and their customers.
If meaningful self-deployment remains practical only for governments, infrastructure companies, and deep-pocketed enterprises, that will tell us something too.
Either way, ML4 is about to put a much clearer price tag on AI sovereignty.
