Mistral and Nvidia Unite LLM and Physical AI for Industry

Mistral and Nvidia clarify their roadmap for AI-assisted engineering. The Mistral Large 4 model, trained on 3,800 Grace Blackwell GPUs, is set to be released as an open-weight model by the end of October. The two partners announce a pipeline connecting design, simulation, and training, along with benchmarks and open datasets.

Detailed Mistral Large 4, Open Release Expected, and a Goal of 1 GW by 2030

Mistral states that Mistral Large 4 has approximately one trillion parameters. This model has been trained on 3,800 Grace Blackwell GPUs in Europe. Its release as an open-weight model is anticipated by the end of October. Mistral also aims to increase its computing capacity to one gigawatt by 2030.

A Common Pipeline to Connect Design, Simulation, and Training

Mistral and Nvidia intend to integrate the Mistral Large 4 language model and Mistral's physical AI with Nvidia's CUDA-X and Warp libraries, as well as Omniverse for 3D simulation. Their goal is to merge design, simulation, and AI training into a seamless process, targeting sectors such as aerospace, automotive, and semiconductors. This partnership aims to enable accelerated engineering simulations and industrial product development through artificial intelligence.

Public Benchmarks, Open Data, and Ongoing Collaborations

Mistral and Nvidia plan to provide public benchmarks and open datasets to evaluate and compare physical models and AI agents used with engineering software. The two companies have already collaborated on Mistral NeMo and are currently involved in creating an open foundation model within the Nemotron Coalition.