CERN and Mistral plan a physics foundation model

CERN and Mistral signed a statement of intent on October 15 to develop foundation models for scientific research. Their initial ambition is a pretrained Large Physics Model for collider data. CERN says it would learn from billions of simulated proton–proton collisions and support reconstruction, particle identification and searches for unusual signals. It could also help evaluate future detector designs.

The announcement reports no trained model, evaluation, access route or discovery from the partnership. CERN's CMS experiment has already tested machine learning for reconstruction, but that was a separate project with a narrower task.

The big change

• What changed: CERN and Mistral agreed to explore a long-term partnership centred on a model intended to work across multiple collider-physics tasks.
• Why it matters: CMS currently reconstructs collision events through specialised steps. A reusable pretrained model could be valuable if it proves accurate across tasks and detector conditions, but the partnership has not demonstrated that result.
• What to watch: The first test will be a documented model and task-specific comparisons with existing methods. The statement gives no release date, validation milestone or public access terms.

From detector signals to physics objects

Particle detectors record traces and energy deposits. Reconstruction software combines those measurements into particles and other objects physicists can analyse. CERN describes today's process as a chain of specialised algorithms. It wants to pretrain one model that could support work across that chain and beyond it.

The announcement specifies simulated proton–proton collisions as the planned training material and names possible downstream tasks. It does not specify the model architecture, the exact input representation, the output interface or how performance would be validated on real detector data. Those details matter when assessing whether one model can serve different experiments or detector designs.

CERN's AI lighthouse plan separately describes a CERN-anchored physics foundation model as shared scientific infrastructure, with first models targeted for 2029. That is a CERN programme target; the October 15 statement does not commit Mistral to that delivery date.

The earlier CMS result answers a narrower question

In a CMS study reported earlier in 2026, researchers tested a machine-learning particle-flow algorithm within the CMS software framework. It takes detector information and produces reconstructed particles. The CMS publication reports performance comparable to its standard particle-flow method, with better jet-energy resolution in specified simulated samples and faster per-event inference under its measured hardware conditions. A small sample of recorded collision data was also used for comparison.

That study shows that machine learning can perform a full reconstruction task under the reported conditions. It predates the CERN–Mistral agreement and is not evidence that the proposed Large Physics Model has been trained, tested or deployed. The planned model has a broader, multi-task aim that will require its own evidence.

CERN calls the proposed system the first foundation model for collider physics. The wording needs care: a February 2025 paper already presented HEP-JEPA under the description “a foundation model for collider physics,” using simulated jet data for pretraining and evaluating downstream classification tasks. That work does not establish the same full-event, multi-task scope CERN now proposes. It does establish that an unqualified priority claim would obscure earlier research.

The partnership joins CERN's collider data and computing expertise with Mistral's model-training work. The partners have yet to publish a model or comparisons with existing tools on defined tasks. BIG CHANGE reviewed the announcement and research records; it has not tested a CERN–Mistral model or reproduced the CMS study.

Sources & further reading

• CERN and Mistral partnership announcement, October 15, 2026. Establishes the statement of intent, planned simulated-collision training and proposed tasks. It reports no model results or access terms.
• CERN AI lighthouse initiatives, read October 15, 2026. Describes CERN's separate physics foundation-model programme and its first-model target. It does not set a Mistral partnership deadline.
• CMS-PFT-25-001, CMS Collaboration, January 24, 2026. Reports the earlier particle-flow reconstruction study, its benchmark conditions and limits.
• HEP-JEPA: A foundation model for collider physics, February 6, 2025. An earlier collider-model paper with a narrower jet-focused evaluation; relevant to the announcement's priority wording.