Level 1 - Absolute Beginner

Mistral AI is a company in Paris. It makes AI models. This week it gave a big model away for free.

The model is called Mistral Large 4. Anyone can download it. People can run it on their own computer.

The model is good at computer security. It can help people protect their systems. But bad people can use it too.

Mistral says the model is safe enough. Some people are not sure.

company

a group of people who work together to sell something

model

a computer program that learns from data

free

costing no money

download

to copy a file from the internet to your computer

security

keeping something safe from harm

protect

to keep something safe

computer

a machine that stores and uses information

safe

not in danger

Level 2 - Elementary

On Thursday October 29, 2026, the Paris based company Mistral AI published a new model called Mistral Large 4. It put the weights online, so anyone can download the model and run it on their own machines.

The model has one trillion parameters. Mistral says it trained the model in European data centres on about 4,000 Nvidia chips over roughly two months. After that it spent three more weeks teaching the model to write code and to defend computer systems.

The licence lets people use the model, change it and build products with it. Companies that want to resell the model as a service must pay for a separate licence. The model is also available on Microsoft Foundry, Hugging Face and OpenRouter.

Mistral says the model is strong at cyber defence, and it reports high test scores. But those scores come from the company itself. A tool that can defend can also attack, and Pierre Stock, a vice president at Mistral, said some bad actors might try to use the technology to attack.

publish

to make something available to the public

weights

the numbers inside an AI model that hold what it has learned

parameter

one of the many values a model uses to make decisions

trillion

a million million

licence

an official permission to use something

resell

to sell something you bought or received from someone else

defend

to keep something safe from attack

data centre

a building full of computers that run software for others

Level 3 - Intermediate

On Thursday October 29, 2026, the Paris based company Mistral AI did something unusual: it published the open weights of its largest model, Mistral Large 4, which means anyone can download the files and run the model on their own machines. The model is natively multimodal and has one trillion parameters, but it uses a mixture of experts design, so only part of it works on any given word. The Hugging Face repository name implies roughly 52 billion active parameters per token, while Mistral's own October preview materials put the figure at 49 billion.

Mistral says it trained the model from scratch in European data centres on about 4,000 Nvidia Grace Blackwell GPUs over roughly two months at around 10 megawatts of power, then spent three further weeks on reinforcement learning that it credits with large gains in coding and in defensive cybersecurity. The company reports a benchmark score of 19.7 percent on Cognition's FrontierCode 2.0, 48 percent on the Artificial Analysis Coding Agent Index, 51 percent on the Artificial Analysis Cyber Index and a 96.6 percent average refusal rate on malicious cyber requests. Every one of those numbers is Mistral's own, and no independent group has checked them.

This is where the argument begins, because a model that is good at finding weaknesses in software is useful to defenders and attackers alike. That is the dual use problem, and this release puts it in the open. Guillaume Lample, the co founder and chief scientist, says the cybersecurity capabilities will enable enterprises and governments to defend themselves, and Booz Allen Hamilton's Brad Medairy says the model matched the vulnerability detection of other frontier cyber models with a compelling cost curve. Pierre Stock, the vice president of science, argues that open weights give users access that cannot be cut off, while conceding that some malicious actors might try to use the technology to attack. Jeff Boudier of Hugging Face describes sovereignty as the freedom to run, customize and build AI on your own terms.

The caution runs in the other direction too. Weights cannot be recalled once people have downloaded them, Mistral itself says it saw potential for real world attack uplift in live traffic and is changing its guardrails, and it is keeping a more capable cyber variant in private preview for vetted partners only. Reports also differed on how much of the model could actually be downloaded on the first day. Since this is the largest open weight model released outside China, and the first from a Western laboratory tuned for cybersecurity work, it is a live test of whether a powerful defensive tool can be handed to everyone without also arming attackers.

open weights

model files that anyone can download and run themselves

mixture of experts

a design in which only part of a model is used for each piece of input

benchmark

a standard test used to compare performance

refusal rate

Level 4 - Advanced

Mistral AI's decision to publish the weights of Mistral Large 4 on Thursday October 29, 2026 reframes an argument rather than settling it. The Paris company released a natively multimodal mixture of experts model with one trillion total parameters, trained from scratch on roughly 4,000 Nvidia Grace Blackwell GPUs in European data centres over about two months at roughly 10 megawatts, followed by three weeks of reinforcement learning that it credits with major gains in coding and defensive cybersecurity. The Hugging Face repository name, 1T-A52B, implies about 52 billion active parameters per token; the company's own October preview materials put the figure at 49 billion. Commercial terms are deliberately permissive, with a modified MIT licence allowing use, modification, fine tuning and commercial deployment, and a separate licence demanded only of firms reselling the model as a service.

What makes the release notable is direction rather than scale. This is the largest open weight model released outside China, and the first time a Western laboratory has shipped downloadable weights explicitly tuned for cybersecurity work. The accompanying numbers are striking and uniformly self reported: 19.7 percent on Cognition's FrontierCode 2.0, which Mistral calls the best result by an open weight model; 48 percent on the Artificial Analysis Coding Agent Index, roughly level with the Chinese open model GLM-5.3 while consuming 47 percent fewer tokens per task; 41 percent on Terminal-Bench 4.0, up thirteen points from the preview; 51 percent on the Artificial Analysis Cyber Index; 84.5 percent on BrowseComp; and 73.3 percent on FinWorkBench. On safety the company claims a 96.6 percent average refusal rate on malicious cyber requests across three public benchmarks and 97.1 percent on the B3 Agent Security Benchmark. None of it has been independently verified, and vendor figures are a claim, not a result.

The dual use asymmetry is the heart of the matter. Guillaume Lample, co founder and chief scientist, says the cyber capabilities will enable enterprises and governments to defend themselves, and Brad Medairy of Booz Allen Hamilton reports that the model matched the vulnerability detection recall of other frontier cyber models with a compelling cost curve. Pierre Stock, vice president of science, defends the open release as access that cannot be cut off by definition because the model is open weighted, while conceding that some malicious actors might try to leverage the technology to attack; on rogue agents he says Mistral does not want to fearmonger and considers the problem solvable. The difficulty is that publication is irrevocable. Defenders at Cisco, Cloudflare, CrowdStrike, Fortinet, Zscaler, Orange Cyberdefense and ENISA gain a cheap and capable assistant, but so does every adversary with a laptop, and proliferation cannot be undone by a later change of policy.

The caveats therefore deserve equal billing with the claims. Mistral itself says it saw potential for real world attack uplift in live traffic and is altering its guardrails in response; the more capable cyber variant stays in private preview for vetted partners only; and accounts differed over how much was genuinely downloadable on day one, with the main checkpoint page reportedly showing an October 31 date while a compressed version shipped at once. Jeff Boudier of Hugging Face frames sovereignty as having the freedom to run, customize, and build AI on your own terms, and chief executive Arthur Mensch insists AI should be a tool for empowerment, not dominance. Both propositions are coherent. Neither answers the question the release actually poses, which is whether a powerful defensive instrument can be handed to everyone without arming the people it was built to stop.