flarephysicsai

Flare, the Large Physics Model CERN and Mistral plan to build

A statement of intent signed on 15 October 2026, a model that learns from simulated collisions, and no date, budget or licence yet.

Updated

On 15 October 2026 CERN and Mistral AI signed a statement of intent for a long-term partnership on foundation models for science (CERN's announcement, 15 October 2026). The first aim is a single pretrained model for collider physics, which CERN calls a Large Physics Model and describes as the first foundation model for its field. Reuters reported it the same day under the name Flare (Reuters, 15 October 2026), and that is the name most coverage now uses. CERN's own release doesn't use it.

Nothing has been built or released yet. A statement of intent sets out what two organisations mean to do together; it isn't a contract, and neither side gave a budget or a timeline.

What it is meant to do

The model would be trained on billions of simulated proton-proton collisions at the Large Hadron Collider, then put to four kinds of work:

  • reconstructing collisions, turning the raw signals in a detector into the particles that made them
  • identifying which kind of particle each one was
  • flagging unexpected signals, collisions that don't look like anything the known physics predicts
  • assessing designs for future detectors

CERN's release points out that today's reconstruction runs as a chain of specialised algorithms, each written and tuned on its own, and that one foundation model could learn across the whole chain. Reuters adds that the model is meant to replace some of those algorithms, and quotes a CERN spokesperson describing it as the idea behind large language models, trained on collision data rather than text. How reconstruction works today.

Known and not known

Everything below comes from CERN's release or the Reuters report. We'll move a row to "known" when CERN or Mistral says more.

QuestionKnown?What we know
Who is building it Known CERN and Mistral AI, under a long-term partnership. (CERN)
What has been signed Known A statement of intent, not a contract with a budget. (CERN)
What it learns from Known Billions of simulated proton-proton collisions at the LHC. (CERN)
What it is for Known Reconstructing collisions, identifying particles, flagging unexpected signals and assessing future detector designs. (CERN)
Its name Known Reuters calls it Flare. CERN's release calls it a Large Physics Model. (Reuters)
Release date Not yet Not announced. Reuters says no timeline was given. (Reuters)
Budget and computing Not yet Not disclosed. (Reuters)
Open weights or licence Not yet Not announced. CERN's CIO says the work should serve the physics community and be shared. (Reuters)
Which experiments and simulations Not yet Not stated. The release says only "LHC" collisions.
Size of the model Not yet Not stated.

What the two sides said

High-energy physics provides a unique environment in which artificial intelligence can be developed and tested

We are proud to work with CERN to explore what large-scale AI models can bring

Johannes Brandstetter, Vice-President of AI for Engineering at Mistral, in the same release

Porcari also told Reuters the partnership should advance AI that can serve the whole physics community and be shared. That is the nearest thing so far to a promise of openness. Mistral's chief executive, Arthur Mensch, wrote in Le Monde the same day that frontier models were now indispensable to scientific progress (as reported by Reuters).

Why CERN, and why Mistral

CERN has the data. It has stored about one exabyte from the LHC, a million terabytes, and expects about ten times more collisions once the machine is upgraded into the High-Luminosity LHC, due around 2030 (CERN). Writing and maintaining a separate algorithm for every step of reconstruction at that scale is the problem a single learned model is meant to ease. The data problem in numbers.

Mistral, based in Paris, builds large models and publishes many of them with open weights. In May 2026 it agreed to buy Emmi AI, an Austrian company working on AI for physics and engineering, and Emmi's co-founder Johannes Brandstetter came with it (Mistral). Days later Mistral announced a physics-AI line of work with industrial partners (Mistral). Who is involved.

"The first foundation model for collider physics"

That is CERN's description, and it needs reading carefully. University groups have published foundation models for particle physics since 2024: OmniJet-α called itself the first cross-task one, and OmniLearned, EveNet and others followed, some trained on more than a billion jets. Most work on single jets rather than whole collisions, and none comes from CERN itself or at the scale a partner like Mistral can train. The collider models before Flare.

Questions we're watching

  • Will the weights be open, and under what licence? Mistral publishes open-weight models; CERN publishes its research openly. Neither has said.
  • Which simulations? CERN's experiments each simulate their own detectors. The release says only "LHC" collisions.
  • How will it be checked? A model trained on simulation can learn the simulation's mistakes. The physics needs proof that it works on real data.
  • Who else can use it? The release says the approach could later serve other sciences.

Papers so far

None. When the first one appears it goes on the papers page the same day, and we'll email everyone who asks below.

Sources