The French AI company Mistral opened the weights of its largest model, Mistral Large 4, to everyone through Hugging Face on Thursday, October 29. The model, which had been offered only as an API preview at the beginning of October, can now be downloaded and run on companies' and governments' own servers.
What happened?
According to data reported by VentureBeat, the model has 1.05 trillion parameters in total, but only 52 billion are active: at each step the model runs only a small fraction of the total. Its context window is 1 million tokens. It can read text and images and generates text. About 3,800 Nvidia Grace Blackwell GPUs were used in its training.
Mistral is positioning the model especially for cybersecurity and coding agents: finding, analyzing and responding to threats in code and IT infrastructure. On the index of the independent evaluator Artificial Analysis, the model scored highest among open-weight models developed outside China. Even so, it still trails Chinese open models such as DeepSeek and GLM. The license isn't fully free: a modified MIT license requires a separate commercial license from companies that sell the model as a service.
Why does it matter?
Today most of the most powerful models can only be used over the internet, as far as the company allows. Open weights mean access can't be cut off later. Pierre Stock, Mistral's vice president of science, told Reuters that once the model is released, restricting access is impossible. That is a big advantage for institutions that don't want to send their data out; but it also means the model could fall into the hands of malicious actors. It's also a concrete step in Europe's "sovereignty in AI" debate.
A link from history
Programs that roam a network on their own, and the idea of stopping them, aren't new. Reaper, written to catch Creeper, which roamed ARPANET in 1971, is regarded as the first antivirus. We told that story in our post Creeper, the first computer "virus", deleted nothing. Today, trillion-parameter models are being written for cyber defense.