Mistral has released the weights of its latest flagship model, Mistral Large 4, for anyone to download, after a preview on 6 October. The company, which calls the model the best open agentic coding model in the world, is presenting it as a tool that organisations can run on their own infrastructure, including for cyber defence.
The model has about 1.05 trillion parameters, of which roughly 52 billion are active on any one task. It offers low, medium and high reasoning modes so users can match depth of reasoning to the job. Mistral says it was trained on 3,800 Grace Blackwell chips in its own European data centres.
The company supplied its own benchmark claims. It says the model uses on average 47 percent fewer tokens per task than a rival Chinese model, GLM-5.3, and that its score on a terminal-based coding test rose from 28 percent to 41 percent. These figures come from Mistral and have not been independently checked.
Access is through Mistral's own interface and through a public model-sharing site where the weights are posted. Beyond coding and security, the company says the model can turn two-dimensional blueprints into three-dimensional design files, analyse financial data, and use satellite images to spot wildfires and monitor crops and power lines.
Pierre Stock, the company's vice president of science, and chief executive Arthur Mensch were quoted in the launch. Mensch said Mistral has raised roughly $6 billion over three years.
Downloadable weights let companies keep sensitive code and data inside their own systems, which is the argument Mistral is making to European customers. How the model performs in outside testing, against closed models from US labs, is the next thing to watch.





