Latest news
Mistral AI put the weights of Mistral Large 4 online on October 29, 2026, and the French lab aimed the release squarely at security teams. The company published the open-weight model as a 1 trillion-parameter, natively multimodal system, and its pitch leans hard on cyber defense instead of general chat.
Reuters reported the release under the headline Mistral AI publicly releases Large 4, targets cybersecurity users. Mistral calls the model “Le Chonk” in its own announcement. It first previewed the system on October 6, so the gap between preview and public weights came to 23 days.
That timing matters. Open-weight releases have become the main battleground between labs that sell hosted access and labs that hand customers the files. Our open-weight model license comparison covered six families before this launch, and Large 4 now adds a new heavyweight to that list. This analysis walks through what Mistral confirmed, what the numbers do and do not prove, and who gains or loses from putting a cyber-capable model in everyone’s hands.
The Shattered Newsletter
The week’s biggest stories in tech, gaming and security, in one email.
Free, once a week. Unsubscribe any time.
Don't miss new tech stories on Google
Add Shattered once in the Google app and our stories show up in your news suggestions.
The release has two parts. Mistral posted the model weights for download on Hugging Face, and it also serves the same model through its own API. Developers who want to run the system on their own hardware can pull the files. Teams that prefer a hosted endpoint can call it the usual way.
Mistral framed the move in plain terms in its announcement: “We release the weights of Le Chonk, advancing the frontier of open models in agentic coding and cybersecurity.” The sentence names two target workloads, and neither is casual conversation. Agentic coding means a model that writes, runs and fixes code across many steps. Cybersecurity means finding, reproducing and patching software flaws.
Mistral also described the model as the best open agentic coding model in the world, according to its official materials. It said enterprises and governments get “frontier AI cyber defence capabilities they can fully own and control.” Those are company claims, and independent reviewers will need weeks to test them. Still, the framing tells you who Mistral wants as customers: security operations teams, public agencies and regulated firms that cannot send sensitive code to a third-party cloud.
One detail remains unconfirmed. Mistral’s cited release materials do not state a price for API access, so any per-token figure circulating today lacks a source. We will update this report if Mistral publishes rates.
The confirmed hardware-agnostic specs are short but significant. Large 4 uses a Mixture-of-Experts architecture, which means only part of the network activates for each token. That design lets a lab claim a trillion parameters without paying the compute bill of a dense trillion-parameter model on every request. It also handles text and other inputs natively, rather than bolting a vision module onto a language model afterward.
A 1M context window changes what a security workflow can look like. A reviewer can feed an entire mid-sized codebase, its dependency manifest and a stack of recent advisories into one prompt. Whether the model keeps its accuracy at that length is a separate question that only independent testing will answer. Mistral has not published a long-context accuracy curve in the materials we reviewed.
The size also sets a practical floor for self-hosting. A model with a trillion parameters needs serious multi-GPU infrastructure even when only a fraction of the experts fire per token. Small teams will most likely use the API, while governments, large banks and telecom carriers can justify the cluster. That split shapes who actually benefits from the open weights in the first year.
Mistral described Large 4 as “one of the world’s strongest AI models for cybersecurity.” It also says the model tells defensive security tasks apart from malicious, offensive ones. That second claim carries most of the weight in the sales pitch, because a model that helps a defender patch a bug can, in principle, help an attacker write an exploit for the same bug.
Mistral co-founder Guillaume Lample made the strategic case in launch coverage. He said the cyber defense capabilities will let enterprises and governments defend themselves against threat actors who are jailbreaking closed models to perform cyberattacks. The argument runs like this: attackers already have access to capable models through jailbreaks, so defenders deserve a model they can run, inspect and tune on their own terms.
The logic has appeal, and it also has gaps. Jailbreaking a hosted model and fine-tuning an open-weight model are different acts. The first fights a vendor’s safeguards in real time, while the second removes them permanently once the files sit on an attacker’s disk. Lample’s point holds only if the defensive benefit outweighs that second risk, and Mistral’s own executives admit the risk exists, as the next sections show.
Based on Mistral’s official framing, three buyer groups stand out. Security operations centers want faster triage and patch drafting. Software vendors want automated bug reproduction before a release ships. Government agencies want a model that never leaves their network. Each group values control over raw leaderboard rank, which explains why Mistral led with ownership language instead of a speed claim.
Mistral reported that Large 4 scores 51% on the Artificial Analysis Cyber Index, a third-party benchmark. Independent indices carry more weight than a lab’s internal tests, so this figure deserves attention. A raw percentage still tells you little without the comparison set, and the fact sheet we hold does not include scores for rival models on the same index.
Readers should treat the number as a snapshot. Benchmarks in this area move quickly, since labs tune for them and new tasks get added. A 51% score means the model passes about half of what the index asks, and the index itself defines what counts as success. Anyone buying on the strength of one score should ask for the task list and run a pilot on their own code.
The 3.4% line in the table is our own subtraction, not a Mistral claim. It assumes the refusal rate is a simple average across the test prompts. Even if that is true, a defender can live with a small leak, while a determined attacker only needs one useful answer.
The second headline number is 82%. Mistral says Large 4 reached that score on a test where the model reproduces a real vulnerability in open-source software and then patches it. Reproduction is the hard part. A model has to read the bug report, build or locate a failing input and prove the flaw triggers, and only then write a fix that removes it without breaking the rest of the program.
That end-to-end shape matters more than a multiple-choice security quiz. Real defenders spend most of their time on exactly this loop. They confirm that a reported flaw is real, they judge its reach and they ship a patch that maintainers will accept. A model that automates even part of it could cut the queue of open advisories that every security team complains about.
Two cautions apply. First, the 82% comes from Mistral’s announcement, and no outside lab has confirmed it yet. Second, benchmark vulnerabilities come from public projects with public fixes, so a model may have seen related material during training. Proprietary code, odd build systems and half-documented internal libraries will test it harder. Our walkthrough on vulnerability management with KEV, EPSS and OSV data shows how much triage work surrounds a single patch, which helps explain why vendors chase automation here.
Mistral reports a 96.6% average refusal rate on malicious cyber requests. In other words, when a prompt asks for something clearly offensive, the model declines in the vast majority of test cases. The company presents this as proof that it can separate defensive help from offensive help.
The figure describes behavior of the model as shipped. Open weights change the situation because anyone with the files can fine-tune the model and strip refusals. Closed vendors keep that lever in their own hands, since they control the endpoint, the logs and the account bans. This difference sits at the center of the policy debate, and regulators are already weighing it. We covered one side of it in Anthropic’s October 15 letter urging courts to block risky AI models.
Closed models face their own leak. Lample’s point about jailbreaks is real, and our report on DeepSeek and Moonshot’s use of GPT-6 Astra showed how capabilities travel even when access is gated. Neither route gives perfect control. The honest framing is that open weights trade vendor-side guardrails for customer-side control, and each buyer must decide which risk they fear more.
Mistral executives did not dodge the downside. Pierre Stock, Mistral’s Vice President of Science, told Reuters: “I’m not comfortable living in a world in which open weights do not exist.” He also acknowledged the flip side of a public release. Once the model is out, he said, “it’s impossible to contain access.”
Those two sentences capture the whole debate. Stock defends openness as a principle, and he concedes that a downloaded file cannot be recalled. Fortune reported a related claim from him after the launch, that Large 4 is the best open-weight model on cyber and coding. Read together, the comments show a company that knows the trade and has decided to accept it, in a Fortune report on the launch that frames the model as a top open option for agentic coding and cyber defense.
Governments are watching the same trade from the other side. Our piece on China’s open-weight AI as weapon or armour lays out how one major power treats open releases as strategy. A Western lab shipping a cyber-capable trillion-parameter model changes the terms of that argument, because the same capability now sits in both camps.
The practical effect is a faster diffusion of skill. Before this week, a mid-tier attacker needed to rent hosted access or jailbreak it. Now the attacker can download a strong model, though running one at this size still takes real hardware and money. That cost barrier filters out hobbyists but not funded groups, and defenders should plan accordingly.
Arthur Mensch, Mistral’s co-founder and CEO, tied the release to a political argument. As SiliconANGLE reported, he said that open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors. For a European lab selling to European institutions, that message lands well. Buyers there have grown wary of depending on a handful of US-based hosted providers.
Euronews framed the launch around the same theme, describing a model released for download that gives companies more control over their cybersecurity. You can read the Euronews account of the release for that angle. Control here has a concrete meaning. A bank can run the model inside its own data center, log every prompt on its own servers and freeze the version it audited, so a vendor update cannot silently change behavior.
Government demand for that kind of isolation is already visible. Sweden recently moved to restrict chatbots in state offices, a step we tracked in Sweden’s block on ChatGPT and Claude in government offices from October 26. A locally hosted model offers one way out of that bind. Officials can keep an AI assistant without sending documents to a foreign cloud.
Security buyers rarely pick on benchmark score alone. They weigh data handling, audit needs, staffing and risk. The table below compares the two delivery models in qualitative terms, using only properties that follow from how each one works. It does not rank specific rival products.
The table shows why the choice is not one-sided. Large 4 wins on control, privacy and tunability. Hosted models win on convenience and on vendor-enforced safeguards. A regulated firm that already runs GPU clusters will lean toward the download, while a ten-person startup will probably use an API from whichever provider fits its budget.
Many large buyers will end up with both. A common pattern is a hosted frontier model for general work and a self-hosted open model for anything that touches secrets, source code or incident data. Large 4 gives that second slot a credible occupant for the first time at this capability tier.
ZDNET’s headline on the launch says Mistral’s Le Chonk is now a top open model for coding, ranking alongside Codex GPT-6.1 Terra. You can see that framing in the ZDNET report on the weights release. The comparison puts an open model in the same sentence as a leading closed coding product, which would have sounded optimistic a year ago.
We hold no verified head-to-head scores, so we will not invent a leaderboard. What we can say is that the open-weight field now has several serious entrants. Our earlier DeepSeek vs Qwen vs Mistral comparison mapped the options before Large 4 landed, and the new release raises the bar for all three families.
The competitive pressure runs in two directions. Open-weight rivals must now answer on cyber tasks specifically, an area many of them skipped in their launch material. Closed vendors must justify their premium to buyers who can see a near-peer they can run themselves. Neither group can ignore a trillion-parameter model with a 1M window and a published cyber score.
Benchmarks will decide how fast that pressure builds. If independent labs confirm the 51% and 82% figures, procurement teams will start adding Large 4 to shortlists. If the numbers shrink under neutral testing, the launch will read as a strong marketing move on top of a decent model.
Mistral previewed Large 4 on October 6, 2026, and shipped the weights 23 days later. A staged rollout like that gives a lab time to gather feedback, harden safeguards and brief partners before the files become permanent public property. It also lets competitors read the preview and prepare responses.
The release also fits a longer pattern. Mistral built its reputation in Europe by shipping open models while US rivals kept their best systems behind APIs. Large 4 pushes that approach to the trillion-parameter scale and adds an explicit cyber focus. Open-weight AI has moved from a hobbyist niche to a core piece of national and corporate strategy within a few years.
Cybersecurity makes the move sharper than a general model release. A chatbot that writes a clumsy poem harms nobody. A model that can reproduce and patch a real vulnerability sits near the same tools that find and weaponize one. That dual-use shadow explains why the same week produced both enthusiastic coverage and pointed questions about containment.
For security vendors, the release lowers the cost of building AI features. A company that once licensed a hosted model to power a triage assistant can now fine-tune a strong open one and keep the margin. That pressure will hit startups that sell thin wrappers over hosted APIs, since customers can now ask why they should pay for a layer that a download replaces.
For buyers, the main gain is leverage. A procurement team can now credibly say it will run an open model if a hosted vendor’s price or terms look poor. Even buyers who never deploy Large 4 benefit from the negotiating position it creates. Security leaders who want a grounding in the basics can start with the shattered.io security section.
Cloud providers face a subtler effect. If large customers self-host, hyperscalers still win GPU rental revenue, but they lose the model margin that sits on top. Expect cloud platforms to list Large 4 in their model catalogs quickly, since refusing to would push customers elsewhere. We could not confirm any such listing at publication time, so treat that as an expectation and not a report.
Mistral itself gets a different payoff. Open weights build trust and adoption, while the API and enterprise support line up the revenue. The company has not published financial details about Large 4, and we make no claims about them.
Several questions remain open at publication time. Mistral’s cited materials give no API price. They do not list rival scores on the Cyber Index in the sources we reviewed. We also found no independent reproduction of the 82% vulnerability result or the 96.6% refusal figure.
Hardware requirements for self-hosting also lack official numbers in our fact base, so we will not guess at GPU counts or memory. The same goes for real-world misuse. No outlet we reviewed documented an attack that used Large 4, and it is far too early for such evidence to exist. Anyone claiming otherwise today is speculating.
These gaps are normal for launch day. Independent evaluators, cloud platforms and security researchers will fill most of them over the coming weeks. We will revisit the story as that data arrives.
These are our analysis, not reported facts, and each could miss.
What is Mistral Large 4?
It is Mistral AI’s 1 trillion-parameter, natively multimodal Mixture-of-Experts model with a 1M context window. Mistral also calls it “Le Chonk.” The company released its weights publicly on October 29, 2026.
Where can I download Mistral Large 4?
The weights are available on Hugging Face. The model is also available through Mistral’s API for teams that prefer not to host it.
How good is Large 4 at cybersecurity?
Mistral reports 51% on the Artificial Analysis Cyber Index and 82% on a test that reproduces and patches a real open-source vulnerability. Independent confirmation of those figures is still pending.
Will the model help attackers?
Mistral says the model separates defensive tasks from offensive ones and reports a 96.6% average refusal rate on malicious cyber requests. Pierre Stock admitted that once weights are public, access cannot be contained, and fine-tuning can change the model’s behavior.
How much does Mistral Large 4 cost?
Mistral’s cited release materials do not confirm a price. Downloading the weights involves hardware and operating costs that depend on your setup.
Who is the model aimed at?
Mistral emphasizes cybersecurity users, including enterprises and governments that want defensive AI they can own and control. Reuters reported the release as targeting cybersecurity users.
When did Mistral first show the model?
Mistral previewed Large 4 on October 6, 2026. The public weights followed on October 29, 23 days later.
The Shattered Newsletter
The week’s biggest stories in tech, gaming and security, in one email.
Free, once a week. Unsubscribe any time.
Marcus Bell covers applied security: authentication, TLS, phishing, and the everyday decisions that keep accounts and data safe. He focuses on turning advice people usually ignore into steps they will actually follow.
Independent explainers on cryptography, security, privacy and the technology behind provably-fair systems. Home of the original SHA-1 collision research.
The Shattered Newsletter: the week’s biggest stories in tech, gaming and security.
About the Author
