Mistral gives away a trillion-parameter model aimed at cybersecurity
An open-weight model is one where anyone can download the model files and run them on their own machines. On Thursday, French company Mistral released the weights of Mistral Large 4, nicknamed "Le Chonk". It was the top AI story on Hacker News, with 1,472 points.
The model is huge: 1 trillion parameters, the numbers inside the model. It uses a Mixture-of-Experts design, so only about 49 billion parameters work on each token (a small piece of text). It is natively multimodal, so it handles more than text. TechCrunch says Mistral targets cybersecurity: companies can run their cyber defence on a model they control, not on a closed one. Smaller community versions were already posted on Hugging Face.
Not everyone is happy about local AI. Sam Altman said running models at home is "a terrible allocation of scarce resources", because it keeps memory chips in personal computers. And the AI Daily Brief says a new benchmark shows a wider gap between open and closed models. On the risk side, Epoch AI published work on the open-weight "patch gap", shared with a headline saying the open model GLM 5.3 can autonomously compromise recent versions of Firefox.
Good news for choice and independence: a top model that is not locked inside a few US companies. And still, a strong open model tuned for cyber work helps attackers as much as defenders, and downloaded weights can never be recalled.
