The letter of intent for a long-term strategic partnership from October 15, 2026 aims at joint foundation models for scientific discovery. Foundation models are large, pretrained machine learning models that can be fine-tuned for different tasks, as described in a review of foundation models in high-energy physics. The centerpiece of the partnership is development of a single pretrained large physics model, which CERN describes as the planned first foundation model for collider physics and which, according to a CERN spokesperson quoted by Reuters, is named Flare, as reported by the portal Numerama.
The partnership brings together CERN's expertise in fundamental physics, scientific data, and large-scale data processing with Mistral's experience in training large models, CERN said. Enrica Porcari, Chief Information Officer of CERN, explained that high-energy physics provides a unique environment to develop and test artificial intelligence with exceptional rigor. The collaboration should advance AI capabilities serving the broader physics community and shareable with it, Porcari said. Johannes Brandstetter, Vice-President of AI for Engineering at Mistral, spoke of the opportunity to jointly push the boundaries of high-energy physics and reach new scientific discoveries.
Why CERN now seeks an AI model for collision data
To date, reconstruction of LHC collisions relies on a chain of specialized algorithms that stepwise convert detector signals into the particles and objects physicists analyze, CERN explains. Machine learning has improved these individual steps for years; a foundation model could now learn across the entire chain and bundle multiple tasks in one model. The industry portal HPCwire describes the step as a transition from multiple specialized AI models to a more general foundation model that could handle multiple scientific tasks. CERN has stored roughly one exabyte of experimental LHC data, equivalent to one million terabytes. With the High-Luminosity LHC, which CERN expects to begin operations around 2030, the number of collisions is to increase tenfold. The accelerator has been shut down since late June 2026 for the so-called Long Shutdown 3, a major program of maintenance, upgrades, and installation work to prepare CERN for high-luminosity operation, as the laboratory announced in June 2026. Phased restart of the accelerator complex is planned for 2028 onward. Jakub Mościcki, head of CERN's storage and data management group, had stated in December 2025 that the stored exabyte represented only 10 percent of what CERN must store and process over the coming decade, as CERN reported.
What Flare is to accomplish
According to CERN, Flare is to bundle multiple tasks in one model: reconstruct collisions, identify particles, flag unexpected signals, and assess future detector designs. Billions of simulated collisions provide training examples where the correct answer is already known by simulation design, as Netics noted on DEV Community. Additionally, HPCwire reported that the new model could reduce the need for separate algorithms, though without guarantee of better results than the existing approach.
Comparison: Earlier foundation models in particle physics
The claim to build the first foundation model for collider physics is CERN's own classification; HPCwire notes that researchers have experimented with similar approaches for several years. In 2024, a team from University of Hamburg presented OmniJet-α and described it as the first cross-task foundation model for particle physics, HPCwire reported. The MLPF foundation model from the study "Machine-learned particle flow as a foundation model for collider physics" was trained with 25 million simulated events, according to detailed information from the authors. Downstream tasks used 350,000 additional training events from a single process, top quark pair production, at a collision energy of 365 gigaelectronvolts, considerably less than the billions of collisions Flare is to be trained on. The authors describe their comparison models in the abstract as models with exclusively kinematic features, though in the methods section they name for these models additional particle identity and, for jet flavor identification, supplementary track features for baseline and latent expanded models equally.
Three gaps between announcement and operational model
According to HPCwire, the partnership is still at an early stage. First, budget, planned computing resources, and a timeline for training start are not public, as AI Polix reported. Second, the announcement contains no benchmark and no comparison with the existing reconstruction chain, as DEV Community reported, and it does not state whether model weights will be openly published, even though Mistral is known for openly accessible models, according to HPCwire. Unclear too is whether a general model can match the accuracy of CERN's existing specialized algorithms, HPCwire noted. Third, the terms for Mistral's access to CERN data are not disclosed, according to DEV Community, even though both partners say the methods could later benefit other sciences, as AI Polix reported. Mistral CEO Arthur Mensch wrote on the day of announcement in Le Monde that leading models have become indispensable for scientific progress; simultaneously he criticized companies that kept their advantages to themselves, as Reuters reported.