CERN, the European particle physics laboratory near Geneva, and the French AI developer Mistral have agreed to work together on foundation models for science. The two organisations signed a statement of intent on October 15, 2026 setting up what they describe as a long-term strategic partnership. Its first goal is a Large Physics Model, which CERN calls the first foundation model built for collider physics and which Reuters reports will be named Flare.
The plan is to pretrain one model on billions of simulated proton-proton collisions from the Large Hadron Collider and then use it for several jobs at once. CERN lists four: reconstructing what happened in a collision, identifying the particles produced, flagging signals nobody expected and helping physicists assess designs for future detectors. A CERN spokesperson told Reuters the approach borrows the idea behind the large language models used in generative AI, except that the training material is collision data instead of text.
That matters because of how LHC data is processed today. Raw detector readings pass through a chain of specialised algorithms, each converting the output of the previous step into objects closer to what physicists analyse, such as tracks, jets and identified particles. Machine learning already improves many of those individual steps, and the spokesperson said CERN has used AI for roughly a decade to turn raw data into pictures of events, but with a separate tool for each stage. A single pretrained model that learns across the whole chain could replace some of those dedicated components, according to Reuters.
Scale is the other driver. CERN says it has stored around one exabyte of experimental data from the collider so far, roughly a million terabytes, and expects the number of collisions to rise tenfold once the High-Luminosity LHC upgrade starts running around 2030. Enrica Porcari, CERN's chief information officer, argued that high-energy physics is an unusually demanding testbed in which AI can be built and checked with great rigour, and said the capabilities developed should benefit and be shared with the wider physics community. Johannes Brandstetter, Mistral's vice-president of AI for engineering, said the collaboration could help push the frontiers of the field.
For Mistral, the agreement gives a commercial model developer a role in one of the largest scientific data environments anywhere, and it fits the company's public stance on science. Reuters notes that chief executive Arthur Mensch wrote in Le Monde on the same day that frontier models have become essential to scientific progress, and took aim at firms that hoard the edge they hold. The company is also preparing the public release of a new open-weight model on October 27.
Several key details are missing. Neither side has published a budget, a schedule, the computing resources involved or a date for training to start. There are no performance results yet, and the access terms for any resulting model, including whether weights will be released openly, have not been specified. The initial work is limited to high-energy physics, although both partners say the methods could later be useful to other sciences.