Physics AI models compared
21 foundation models trained on physics, plus a dataset and a toolkit people often ask about: what each learns from, how big it is and whether you can download it.
Updated 17 October 2026
Flare would join a fast-growing group of models trained on physics data instead of text. Each row comes from the
model's own paper, model card or code, checked on 17 October 2026. Where a source doesn't give a number we say "Not
stated" rather than guess. A code licence is not always the licence for the weights, and we say which we mean.
Collider physics
Model Who Learns from Size Open weights First out
Flare (Large Physics Model) Planned CERN and Mistral AI
Billions of simulated proton-proton collisions at the LHC (planned)
Not stated
Not announced
Announced Oct 2026
Particle Transformer (ParT) Huilin Qu, Congqiao Li, Sitian Qian
JetClass, 100 million simulated jets of 10 types
2.1M
Not stated, MIT (code)
Feb 2022
Masked Particle Modeling Tobias Golling, Lukas Heinrich, Michael Kagan and others
JetClass, self-supervised
Not stated
Not stated, MIT (code)
Jan 2024
OmniJet-α Joschka Birk, Anna Hallin, Gregor Kasieczka (University of Hamburg)
JetClass jets turned into tokens; a follow-up used 178 million real CMS jets
Not stated
Yes
Mar 2024
OmniLearn and OmniLearned Vinicius Mikuni, Benjamin Nachman, with Wahid Bhimji and Chris Harris (Berkeley Lab, NERSC)
More than a billion jets, including ATLAS and CMS open data, JetClass and H1
Not stated
Yes
Apr 2024
HEP-JEPA Jai Bardhan and others
JetClass, 100 million jets
Not stated
Not stated
Feb 2025
EveNet Ting-Hsiang Hsu, Bai-Hong Zhou, Qibin Liu and others
500 million simulated collision events
Not stated
Not stated
Jan 2026
Machine-learned particle flow (MLPF) Joosep Pata, Farouk Mokhtar, Michael Kagan and others (CMS)
Simulated detector events
Not stated
Not stated
Jan 2021
Jet scaling-law models Jan-Lucas Uslu, Benjamin Nachman, Christopher Ré
Collider jets
Several sizes
Yes
Jul 2026
NEXUS Liangyu Wu, Qibin Liu, Alexander Yue and others
LHC charged-particle track features
3M
Not stated
Jul 2026
Prompting Particle Physics Nilotpal Kakati, Daniel Murnane, Baran Hashemi and others
Simulated jets as tracks, calorimeter cells, clusters and particles
Not stated
Not stated
Sept 2026
Simulation and fluids
Model Who Learns from Size Open weights First out
Multiple Physics Pretraining Michael McCabe, Bruno Régaldo-Saint Blancard, Liam Parker and others (Polymathic AI)
Fluid simulations of several physical systems
7.6M to 116M
Yes, MIT (code)
Oct 2023
The Well Dataset Ruben Ohana, Michael McCabe, Lucas Meyer and others (Polymathic AI)
15 TB of simulations in 16 datasets: fluids, supernovae, biology, acoustics
Not stated
Not a model, BSD-3-Clause (code)
Nov 2024
Walrus Michael McCabe, Payel Mukhopadhyay, Tanya Marwah and others (Polymathic AI)
19 kinds of physical scenario from fluids to plasma and astrophysics
1.3B
Yes, MIT
Nov 2025
NVIDIA PhysicsNeMo Toolkit NVIDIA
Not a trained model
Not stated
Not a model, Apache-2.0
Jan 2023
Astronomy
Model Who Learns from Size Open weights First out
AION-1 Liam Parker, François Lanusse, Jeff Shen and others (Polymathic AI)
More than 200 million observations from five sky surveys, including SDSS, DESI and Gaia
300M to 3.1B
Some, MIT
Oct 2025
Weather and climate
Model Who Learns from Size Open weights First out
Aurora Microsoft Research AI for Science, with Cambridge, Amsterdam, Penn and others
More than a million hours of Earth-system data
1.3B
Yes, MIT
May 2024
GraphCast Google DeepMind
40 years of ERA5 reanalysis
36.7M
Yes, CC BY 4.0 (weights)
Dec 2022
GenCast Google DeepMind
40 years of ERA5 reanalysis
Not stated
Yes, CC BY 4.0 (weights)
Dec 2023
AIFS ECMWF
ERA5 and operational analyses
Not stated
Yes, CC BY 4.0
Jun 2024
Prithvi WxC IBM and NASA
160 variables from NASA's MERRA-2 reanalysis
2.3B
Yes, CDLA-Permissive-2.0
Sept 2024
The Sun
Model Who Learns from Size Open weights First out
Surya NASA and IBM
Nine years (about 218 TB) of Solar Dynamics Observatory images
366M
Yes, Apache-2.0
Aug 2025
Materials and chemistry
Model Who Learns from Size Open weights First out
MatterGen Microsoft
Crystal structures from the Materials Project and Alexandria
46.8M
Yes, MIT
Dec 2023
UMA Meta FAIR Chemistry
About half a billion 3D atomic structures, including the OMol25 dataset
1.4B (about 50M active)
On request, FAIR Chemistry License v1
Jun 2025
What the table shows
Most of these models are open: weather, the Sun, astronomy and materials all have large models anyone can download.
Collider physics is the odd one out. Its models come mostly from university groups, few say how big they are, and
no LHC experiment has released one as a foundation model. That is the gap CERN and Mistral say they want to fill.
Before Flare, the collider models in more detail: four years of collider
foundation models .