CERN and Mistral Plan a "Large Physics Model" for Collider Data

CERN and Mistral have signed a statement of intent to build a long-term partnership on foundation models for scientific discovery. Its first project is an ambitious one: a single pretrained Large Physics Model, which CERN calls the first foundation model for collider physics, trained on billions of simulated proton-proton collisions at the Large Hadron Collider (LHC). The details come from the joint announcement as published by Mirage News and a Reuters report carried by bdnews24.

It is worth being precise about what this is. There is no model, no weights, no benchmark and, per Reuters, "no timeline or budget was disclosed." This is a partnership announcement with a stated technical goal. That makes it a story about direction, not about a release, and the useful work is understanding the problem and what to watch for.

TL;DR

The problem: a long chain of specialised algorithms

When protons collide in the LHC, detectors record electrical signals. Turning those signals into the particles and other objects physicists analyse is called reconstruction, and today it relies on, in CERN's words, "a chain of specialised algorithms that progressively transform detector signals into the particles and other objects used for physics analysis."

Machine learning already plays a large part in that chain, improving performance across a range of specialised tasks. The announcement argues that foundation models offer the potential to build on these advances by "learning across the whole reconstruction chain and supporting multiple tasks within a single model." Reuters summarises the aim as replacing some of the specialised algorithms now used to reconstruct and identify particles; the joint statement is more cautious, speaking of building on existing advances.

Why simulated collisions?

The model is to be trained on simulated collisions, not on the experimental archive. CERN says it holds around one exabyte (one million terabytes) of experimental LHC data. Simulation gives labelled ground truth: in a simulated event you know which particles were produced, so you can train and measure how well a model reconstructs them. Real data has no such labels, and physicists must instead validate models by comparing simulation against measurement.

This design also raises the standard open question for physics foundation models: simulation mismatch. A model that learns the simulator's quirks may behave differently on real detector data. Expect the eventual validation to hinge on how well the model transfers from simulation to measurement, and on how it handles calibration and systematic uncertainties, because a physics result has to quantify its errors, not just maximise accuracy. CERN's quote frames the field as one where AI can be "developed and tested with exceptional rigour."

Why this matters for the scale of data

The announcement notes the number of collisions is set to increase tenfold with the High-Luminosity LHC, expected to begin operation around 2030. More collisions mean more computing for simulation and reconstruction, and more pressure to find faster, unified methods. A single model that handles multiple tasks could reduce the engineering cost of maintaining many separate algorithms, though that is a hope rather than a demonstrated result.

Who is involved, and why Mistral

CERN's Chief Information Officer Enrica Porcari said: "We are partnering with Mistral, a leading European AI company, to advance AI capabilities that can serve the whole physics community and be shared." The word "shared" is a signal worth tracking: whether the resulting model, code or datasets are released openly is the biggest unanswered question, and nothing in the announcement commits to it.

Mistral's side was voiced by Johannes Brandstetter, Vice-President of AI for Engineering: "We are proud to work with CERN to explore what large-scale AI models can bring to the growing complexity of particle physics." The statement's language, "explore" and "have the potential," is conditional throughout.

Mistral has been building out science and engineering capabilities. Its own news page lists a physics AI introduction and a recent time-series release, Timestral, alongside its frontier work such as Mistral Large 4. Foundation models for numerical, non-text data are becoming a second front for the lab, separate from chat and coding assistants.

How it fits into AI for science

Collider physics joins a growing list of domains where labs are training large models on scientific data rather than text: weather, genomics, materials and more. For comparison, see Microsoft's open-source bias correction for 2-to-6-week weather forecasts, Ai2's HIRO ACE v2 extreme weather emulator, and Nvidia's European AI-in-science pledge. It also sits beside Europe's sovereignty push, which we cover in our explainer on sovereign AI: the announcement itself stresses "European excellence in AI and fundamental science."

Is it really the first? Prior work in high-energy physics

CERN describes the planned model as the first foundation model for collider physics. That framing is worth checking against the literature, because physicists have been exploring the idea for some time. A 2025 arXiv paper titled Foundation models for high-energy physics surveys the approach, and public work includes OmniJet-alpha, a cross-task foundation model for particle physics presented at a CERN machine-learning working group, and Aspen Open Jets, a dataset of about 178 million jets derived from CMS 2016 open data and used to pretrain OmniJet-alpha. Those projects mostly focus on jets or on specific tasks, so "first" likely refers to the scope CERN and Mistral describe: one pretrained model across the whole reconstruction chain. That scope is the claim to test when results appear.

What a "foundation model for detector data" means in practice

A language model reads tokens; a collider model must read something else. Detector output is a sparse, irregular set of hits, energy deposits and tracks, with no natural ordering, and an event can contain thousands of such objects. A model for this data has to be invariant to the order of the inputs and respect symmetries of the detector geometry, so architecture choices differ from a text transformer. The announcement does not say which design Mistral and CERN will use, and we are not guessing at one.

What it does say is that one model would serve many tasks: reconstruction, particle identification, anomaly flagging and detector design evaluation. The anomaly-flagging item is notable. A model pretrained on simulation that flags "unexpected signals" is, in effect, an outlier detector against the Standard Model-based simulator, and physicists will want to know how it separates genuine new physics from simulator imperfections or detector faults. The detector-design item is also distinctive: if a model is cheap enough to run, designers of future detectors could test layouts in simulation before building hardware.

Questions physicists will likely ask

• Can the model quantify uncertainty? Results in physics are reported with systematic uncertainties, so a black-box output is not enough.
• Is it validated on real data? Training on simulation is convenient, but the real test is performance on measured collisions.
• Who can use it? Porcari's phrase about capabilities that "can serve the whole physics community and be shared" suggests wide access is intended, but no licence has been announced.
• What does it cost to run? Trigger systems and reconstruction run at enormous rates, so inference cost matters as much as accuracy.

What is claimed versus what is verified

What to watch next

1. A technical paper or preprint describing architecture, tokenisation of detector data and training scale.
2. Open release terms: weights, code, and the simulated datasets, given CERN's "shared" framing.
3. Validation: comparisons against the existing reconstruction chain on simulation and on real data, including systematic uncertainties.
4. Compute: where it is trained, given the scale implied by "billions" of collisions.
5. Experiment buy-in: LHC experiments each have their own software and collaboration processes, so adoption is not automatic.

What this means if you build with AI

The practical takeaway is not a tool to download today. It is a signal that foundation-model pretraining is moving to structured scientific data, where the value lies in handling many tasks from one representation. If you work on simulation-heavy domains, the pattern to watch is the same: pretrain on large synthetic data with known ground truth, then validate carefully against real measurements. Mistral's Timestral announcement used the same synthetic-first idea for time series, with real data added in a second training stage.

Related reading

• Timestral: Mistral's time-series foundation models
• Mistral Large 4 open-weights preview
• Microsoft PBC weather bias correction
• HIRO ACE v2 weather emulator
• Nvidia's Europe AI-in-science pledge
• What is sovereign AI?

Details are accurate as of October 15, 2026, based on the joint CERN and Mistral statement and Reuters; no model or results had been announced.