CERN and Mistral name a physics model, but not the experiments that own the data
Their statement of intent names no funding, compute or LHC experiment. Under CERN's policy, each experiment decides what data can be released.
CERN says it will build the first foundation model for collider physics. That means one large AI model, trained once, that can be reused for many tasks. If it works, one trained system could replace a long chain of hand-built algorithms. Today that chain turns detector signals into particles.
The project depends on the LHC experiments that hold the data. They are not parties to the statement, so the announcement has not yet become a project.
In its 15 October release, CERN says it and Mistral signed "a statement of intent to establish a long term strategic partnership." The partnership is meant to develop "foundation models for scientific discovery." A statement of intent is a promise to negotiate. CERN has long run similar arrangements with industry partners through its openlab framework.
The release describes "leveraging these combined capabilities and resources" but quantifies none of them. It states no sum of money, amount of computing power, timeline or licence for the finished model. No LHC experiment is named.
The one CERN official quoted is Chief Information Officer Enrica Porcari. She speaks of advancing AI capabilities "that can serve the whole physics community and be shared." That describes the intended use and leaves open who pays to build the model.
The technical plan is more specific. CERN says it will develop "a single pretrained Large Physics Model" trained on billions of simulated proton proton collisions at the Large Hadron Collider. It is meant to support tasks from reconstructing collisions and identifying particles to flagging unexpected signals and evaluating future detector designs.
Reconstruction is the part a foundation model would change most. Today it "relies on a chain of specialised algorithms that progressively transform detector signals into the particles and other objects used for physics analysis." A foundation model would learn across that whole chain in one system.
The release also cites around one exabyte of experimental data stored to date, which is one million terabytes. It cites a tenfold rise in collisions with the High Luminosity LHC upgrade, expected to begin operation around 2030. That is the scale of the archive and the future load.
The experiments decide what leaves their archives. CERN's open data policy for the LHC experiments sets the terms. A working group of CERN management, the experiments and CERN IT wrote it in December 2020. ATLAS, CMS, LHCb, ALICE and TOTEM each hold their data under their own policies.
Under that policy, reconstructed data follow "an appropriate latency period", with releases aimed at "within five years of the conclusion of the run period" under a CC0 licence. An experiment may hold data back "if there are active analyses ongoing." Raw data are tighter still, since "general direct access to the raw data is not even available to individuals within the collaboration."
One clause points toward Mistral. Representative subsets of raw data "useful for example for studies in the machine learning domain" can be released "at the discretion of each experiment." Every route to the data runs through an experiment's own choice. The policy also provides education and outreach subsets. Those are "not intended nor adequate for the publication of scientific results," so they will not train a research-grade model.
The model would also need to be checked against real detector data. Simulation can teach a model the patterns, but physicists will want proof that it behaves on recorded collisions. That proof requires the experiments to agree to a test and share its results.
The policy governs public release. It does not address what the experiments would let CERN IT or a commercial partner use internally to train a model. No document cited in the release settles it, so the decision rests with the collaboration boards.
The real negotiation is internal. The spokespersons and boards of the experiments, not the signatories, will set the terms.
A model trained on unreleased internal simulation and reconstruction could beat any public one. It would also raise questions of credit for physics results that the experiments guard closely.
The release gives the model's purpose in plain terms. It would help physicists flag unexpected signals and evaluate future detector designs. Both uses need the experiments' trust.
Mistral sells models to enterprises and public institutions. An agreement with CERN IT alone would give it a reference customer and a name. Access to experiment simulation would give it the training material that makes the model worth having. The statement of intent covers only the first.
CERN brings central reconstruction, computing scale and a data volume that the High Luminosity upgrade will multiply. Mistral, which describes itself as a global AI lab building open-weight models, gains a flagship science reference. The release does not say the Large Physics Model will be open-weight, and Porcari's word "shared" is the only signal on the point.
Any funder judging the partnership faces the physicists' question: what may the model see, and who may use it afterwards?
A signed agreement by the end of 2027 would turn this statement into a project. It would need to name the participating experiments, the computing and the licence for the model. If the only new paper by then is another statement, the model will remain an intention. The High Luminosity LHC start around 2030 will set the real deadline.