Mistral released the 1 trillion-parameter Mistral Large 4 model

Mistral has introduced Mistral Large 4, a massive model featuring 1 trillion parameters. It is currently available for developers through a public preview on Mistral Studio.

Built in Europe, this model is designed to handle complex agentic workflows, meaning it can use tools and plan steps to finish tasks. The team focused heavily on professional sectors, including legal analysis, finance, and scientific research. It reportedly excels at coding and math, and can manage tasks like generating chemistry simulations.

One significant feature is its ability to process and ground images. It can inspect technical drawings or analyze geospatial data, such as satellite imagery, to help users find specific details. In internal benchmarks, it performed well at understanding complex charts and visual information.

For security professionals, Mistral emphasizes that the model can assist with vulnerability research without the rigid refusals seen in some other closed systems. It was trained to be robust against prompt injection attacks, which are attempts to trick the AI into ignoring its safety rules.

Mistral intends to release the model weights by the end of October. This move is intended to give organizations more control over how they run their security operations.

While the preview is live, the company is still conducting red-teaming exercises with partners and authorities. It is important to note that full model weights are not public yet, and performance on certain specialized benchmarks may vary depending on the specific environment and deployment method.