Independent tracker · not CERN, not Mistral
Tracking Flare, the physics AI from CERN and Mistral
On 15 October 2026 CERN and Mistral AI agreed to build a foundation model for collider physics, a Large Physics Model that learns from billions of simulated LHC collisions instead of text. We follow it from announcement to first paper, and keep what is known apart from what isn't.
- Status
- Statement of intent, signed 15 October 2026
- Learns from
- Billions of simulated LHC collisions
- Papers
- None yet
- Release date
- Not announced
- Open weights
- Not announced
Checked 17 October 2026 against CERN's release and Mistral's news page.
A model trained on collisions, not words
Large language models learn by reading text. Flare would learn the same way from particle collisions: billions of them, simulated at CERN's Large Hadron Collider, until it has a general sense of what a collision looks like inside a detector. Then it gets tuned for each job.
CERN names four jobs: reconstructing collisions, identifying particles, flagging unexpected signals, and assessing designs for future detectors (CERN's release). Today each of those uses its own chain of specialised algorithms.
Today
- Detector hits
- Track finding
- Vertices
- Calorimeter clusters
- Particle flow
- Particle ID
- Jets and physics objects
Each step is its own algorithm, tuned by its own team.
The plan
- Detector hits
- One pretrained model, learned from billions of simulated collisions
- Reconstruction
- Particle identification
- Unexpected signals
- Detector design studies
The same model tuned for each job.
Latest on the timeline
-
· Flare
Reuters reports the model as "Flare"
Reuters' headline names it Flare and quotes a CERN spokesperson: the same idea as the large language models behind generative AI, trained on collision data instead of text.
Source: Reuters (via KSL)
-
· Flare
CERN and Mistral sign a statement of intent
A long-term partnership whose first aim is a pretrained Large Physics Model for collider physics, trained on billions of simulated proton-proton collisions. No budget, timeline or licence is given.
Source: CERN
-
A "prompting" model maps any reconstruction step to any other
One tokenised model takes any mix of tracks, calorimeter cells and particles and predicts the others. It is the closest published idea to learning the whole reconstruction chain at once.
Source: arXiv 2609.31862
-
The LHC stops for its long upgrade
Run 3 physics was planned to end on 29 June 2026, starting Long Shutdown 3, the works that turn the LHC into the High-Luminosity LHC.
Source: CERN
Why CERN wants it
- 1 EBof LHC data stored so far, a million terabytes
- 40Mbunch crossings a second inside the LHC
- ~10×the LHC's design collisions, once it becomes the High-Luminosity LHC
- 2030when the upgraded machine is due to start
Sources: CERN, CMS, HL-LHC media kit, CERN's schedule. The data problem in full
Read on
What was announced
The statement of intent, the quotes, and what it left open.
Papers and code
No Flare paper yet. The collider foundation models it builds on, with links.
Physics AI models compared
23 models and tools from jets to weather to crystals: who made them and what's open.
Before Flare
OmniJet-α, OmniLearned, EveNet and the rest: four years of collider foundation models.
Learning from simulation
Why train on fake collisions, and what can go wrong when real data differs.
Who is involved
The people named so far at CERN and Mistral, and the physicists whose work it follows.
Hear the day the first paper lands
Nothing about Flare is public yet beyond the announcement. Leave your email and we'll write once, when the first paper, model or date appears, with what it says. Then we delete your address.