The collider foundation models before Flare
Flare is not the first attempt at a foundation model for particle physics. It may be the first from CERN itself, and the largest.
The idea of pretraining one model on collision data and tuning it for many jobs has been around since about 2022, when the JetClass dataset of 100 million simulated jets gave groups something big enough to pretrain on. Since then university groups have tried several recipes. Most work on jets, the sprays of particles from a single quark or gluon, rather than on whole collisions.
In order of first publication:
Machine-learned particle flow (MLPF)
Joosep Pata, Farouk Mokhtar, Michael Kagan and others (CMS) · Jan 2021
Rebuilds the particles in a collision from raw tracks and calorimeter hits in one learned step, replacing hand-written particle-flow rules. Proof inside CMS that one trained model can do a large share of reconstruction. A June 2026 paper argues MLPF already works as a foundation model.
Machine-learned particle flow (MLPF) in detail
Particle Transformer (ParT)
Huilin Qu, Congqiao Li, Sitian Qian · Feb 2022
A transformer that tags jets by reading the particles inside them, released with the JetClass dataset it was trained on. Not a foundation model itself, but JetClass became the standard pretraining set for the jet models that followed, and CMS uses descendants of this design in its taggers.
Particle Transformer (ParT) in detail
Masked Particle Modeling
Tobias Golling, Lukas Heinrich, Michael Kagan and others · Jan 2024
Hides a particle in a jet and trains the model to predict it, the particle-physics version of the masked-word trick behind BERT. One of the first serious attempts at self-supervised pretraining on collision data, the step a model like Flare would rely on.
Masked Particle Modeling in detail
OmniJet-α
Joschka Birk, Anna Hallin, Gregor Kasieczka (University of Hamburg) · Mar 2024
Turns jets into a vocabulary of tokens, learns to generate them like a language model, then transfers that knowledge to classifying jets. Its paper claimed "the first cross-task foundation model for particle physics" in 2024, so CERN's "first foundation model for collider physics" is best read as the first at this scale and for whole collisions.
OmniLearn and OmniLearned
Vinicius Mikuni, Benjamin Nachman, with Wahid Bhimji and Chris Harris (Berkeley Lab, NERSC) · Apr 2024
A framework of small, medium and large jet models meant to serve every jet task, from top tagging to anomaly detection in CMS data. Trained on more than a billion real and simulated jets, with its checkpoints public. Flare is meant to go further, from jets to whole collision events.
OmniLearn and OmniLearned in detail
HEP-JEPA
Jai Bardhan and others · Feb 2025
Pretrains by predicting the hidden parts of a jet in an abstract representation instead of reconstructing the particles themselves. Shows there is more than one way to pretrain on collisions. CERN has not said which approach Flare will take.
EveNet
Ting-Hsiang Hsu, Bai-Hong Zhou, Qibin Liu and others · Jan 2026
An event-level model that mixes self-supervised learning with physics labels; its authors use it to rediscover the Upsilon meson in CMS open data. The closest earlier model to what CERN describes: pretrained on whole simulated events rather than single jets.
Jet scaling-law models
Jan-Lucas Uslu, Benjamin Nachman, Christopher Ré · Jul 2026
Shows the loss of a large jet model can be predicted from small ones to within 1%, and releases five pretrained models. Scaling laws are how a lab decides how big to build. This is the first published evidence that they hold for particle-physics models.
Jet scaling-law models in detail
NEXUS
Liangyu Wu, Qibin Liu, Alexander Yue and others · Jul 2026
A small autoencoder model built to run in real time, which its authors also adapt to gravitational waves and flood forecasting. The opposite end from Flare: tiny and fast, for triggers and edge hardware rather than offline reconstruction.
Prompting Particle Physics
Nilotpal Kakati, Daniel Murnane, Baran Hashemi and others · Sept 2026
One tokenised model that takes any subset of reconstruction objects and predicts any other, covering particle flow and detector simulation. The nearest published idea to CERN's aim of a model that learns across the whole reconstruction chain at once.
Prompting Particle Physics in detail
Where that leaves Flare
Three things set the CERN and Mistral plan apart from all of these, if it happens as described: it is meant to cover whole collisions and the whole reconstruction chain, not one task; it comes from the laboratory that runs the detectors; and it has a partner that trains models with hundreds of billions of parameters. Everything else, size, recipe, licence, is still to be announced. What is known so far.