Mistral, the French AI company, has released the weights of the finished Le Chonk, its trillion-parameter model. That makes the model a file that anyone can download and run on their own hardware. It is tuned for writing code and for security work: finding a flaw in software, proving it is real, and fixing it.
A model you hold cannot be taken away
A model you rent through a company's website works only while that company allows it. It can refuse a request, change its price or be switched off for your account. A model you have downloaded behaves like software on your own server, and nobody can reach in and change what it will do.
That matters most on the worst day. When Hugging Face was attacked in July, the refusal from the commercial services was one problem. The other was that the attack records referred to company credentials the attacker had reached, and sending those records to an outside service would have sent that sensitive material outside the company. Running the model on its own servers solved both at once.
Most businesses cannot run a trillion-parameter model themselves, because it is far too big for office computers. They will meet it inside their security provider's tools instead, and Elastic, a maker of security software, already supports Mistral's open models.
Mistral calls Le Chonk the best open model for security, and that claim is only partly true. On a coding-agent test it ties the best Chinese open model while using 47% fewer tokens, which means it costs less to run. On its own security chart, a few Chinese models score close to it or above it. The ranking matters less than the fact that a model this capable can now run inside a company with no foreign provider in the loop.
The same file works for the attacker
Mistral trained Le Chonk to refuse clearly malicious cyber requests, and in its own tests it refuses 96.6% of them. That holds for the model as shipped. A refusal is a learned habit stored in the weights, and whoever downloads the weights can train a new habit over it. The refusal protects the hosted version, and it does not protect a copy someone has modified.
The attackers' side is already speeding up. Free Chinese models like GLM-5.3 could already build working exploits, so Le Chonk does not open a door that was shut. Le Chonk adds a European-built one, and once a model is downloaded, no company's policy can take it back.
In October, a researcher using an AI model found serious flaws in PHP, a programming language that websites around the world run on. One person did in about a week what would have taken a skilled team months. Sites went offline while operators patched, and millions of attempts to exploit the flaw followed the announcement. Expect the gap between a flaw becoming public and criminals using it to keep shrinking.
What a small company should check
Take a 40-person accounting firm that pays a provider to watch its network. One morning an employee's account is hijacked, and the provider's analysts need an AI to read the login records. If the provider uses a hosted service with strict safety rules, the service may refuse the records because they look like attack material, and the analysts lose hours. If the provider runs an open model on its own servers, the records get analyzed and stay in-house. A refusal at that moment pushes the work back toward the days it took before AI could help.
So ask your security provider which AI reads your logs, whether it will analyze real attack data, and whether it runs on their own servers. Then turn on automatic updates wherever you can, because the time you have to fix a flaw is getting shorter.