BMW Group and Mistral AI announce a partnership focused on collision simulation. The objective is to improve both the quality and the speed of this engineering task, which today consumes a great deal of computing resources.
Each week, BMW runs thousands of virtual collision simulations. These simulations generate vast amounts of technical data. Over time, the automaker has built a historical database exceeding one petabyte, dedicated to this purpose. This corpus records the structures of vehicles and the behavior of materials under impact conditions.
It is on this foundation that BMW and Mistral AI train a specialized AI model, distinct from general-purpose models.
Industrial models rather than generalist
BMW relies on what it calls “Large Industry Models” (LIM). These models embed domain-specific knowledge directly into their architecture. They require industrial data, deep domain expertise, and tailored technical environments. Their training is conducted from the company’s real development processes.
This approach differs from a traditional use of generative AI. It isn’t about adapting a consumer-facing model to an industrial context, but about building a model from proprietary and specialized data.
BMW presents this partnership as a first step. The group aims to extend this approach to other areas of automotive development and across its entire value chain. Collision simulation thus serves as a testing ground before broader deployment.
