French AI company Mistral AI on Thursday released the weights of Mistral Large 4, a model with about 1.05 trillion parameters, about three weeks after a public preview. Companies and governments can download it from Hugging Face, customize it and run it on their own infrastructure.
What the model can do
It is a mixture-of-experts model that activates only about 52 billion parameters per task, with a one-million-token context window. According to Mistral, it narrowly leads all open models on Artificial Analysis' coding agent index and uses fewer tokens than rivals. It was trained on 4,000 Nvidia GPUs in Mistral's data centers in Europe.
What the testers say
Independent tests are less clear-cut. According to Artificial Analysis, it is the best open model developed outside China, but still trails the Chinese models DeepSeek V4.1 Flash and GLM-5.3, mainly in coding. There is also a practical limit: running a model this size requires expensive GPU infrastructure that most small companies cannot afford.
The cyber angle
Mistral explicitly aims the model at cyber defense, and says it refuses 96.6% of malicious requests. Even so, the company admitted that during the preview it saw in real use "a potential for uplift to conduct real attacks", and it is tightening safeguards. That is exactly the dilemma of open models: the same tool that helps defenders is available to attackers.
