Meta Opens Its Muse Models to Five Universities for Medical Research

Meta is giving researchers at five US universities access to its latest AI models through a program focused on scientific discovery, medical training, and clinical research. The program covers Caltech, Ohio State College of Medicine, LSU, Indiana University, and UTHealth Houston. Support includes model access, collaboration with Meta researchers, and funding for selected projects.

The medical track covers patient outcomes, clinical education, and care in low-resource settings. The fundamental-science track will test whether models can formulate and prove conjectures, predict unseen outcomes, construct verifiable artifacts, and generate hypotheses for researchers to evaluate.

A model stack built for local use

Open-weight models expose their learned parameters, allowing institutions to run inference on their own hardware under the applicable license. Closed-weight models keep those parameters private and provide access through an API. The program spans both approaches:

• Muse Wonder and Muse Spark: closed-weight models available through Meta’s first-party API.
• Muse Spark Open: Meta describes it as the company’s most capable open-weight model.
• Muse Glimmer: a 30-billion-parameter open-weight model that Meta says can run on a single consumer GPU.
• Computer vision: SAM for image segmentation, DINO for self-supervised image features, and V-JEPA for learning and predicting how visual scenes change.
• SeamlessM4T: multilingual speech and text translation.
• Wearable hardware: Meta Glasses and VR headsets for selected projects.

Caltech demands checkable answers

Across five Caltech labs, researchers will test models on problems with objective or independently verifiable results. Mathematician Tony Yue Yu is developing MathDB Frontier, a community-curated benchmark based on MathDB, which contains more than 86,900 research-level open problems. Submissions must include machine-checkable proofs, giving evaluators a concrete correctness test.

Chemist Sarah Reisman is building a multi-product reaction benchmark from unpublished electronic lab notebooks, reducing the chance that answers appeared in model training data. Anima Anandkumar is developing TorchLean, a framework that connects AI agents to the Lean 4 proof assistant so their code and proofs can be formally verified. Matt Thomson is constructing an “AI theorist” that writes and tests biological models based on partial differential equations, which describe how systems change across space and time.

Patient data stays inside each institution

The medical projects emphasize open-weight models that participating universities can run within their own computing environments. The universities will conduct the research independently and retain control of patient and clinical records. Meta says the data will not be shared with, processed by, or otherwise made available to the company.

• Ohio State, heart failure: A multimodal system will combine abdominal CT scans, longitudinal health records, and electrocardiogram data to identify early risk of heart failure with preserved ejection fraction, or HFpEF. The condition accounts for roughly half of heart failure cases and can be difficult to detect before symptoms develop.
• Ohio State, radiology: Thoracic radiologists at Ohio State interpret about 200 chest X-rays per day, while workloads at some institutions reach 1,000. A Muse-based system will draft reports from images, patient histories, and previous studies. A related study will test statistical methods for measuring and correcting bias when one language model grades another.
• LSU, infant tracheostomy: The Eyes-On project will use SAM-based video tracking on bedside cameras to detect displaced tracheostomy tubes in infants.
• Indiana, newborn care: NeoSmartML+ will run SeamlessM4T, NLLB-200, and SAM 3 offline on low-cost Android devices to support Kangaroo Mother Care in areas with limited connectivity.
• UTHealth Houston, glioblastoma: ClinAgent GBM will use Muse Glimmer to coordinate analysis across multimodal electronic health records. Initial testing covers 65 patient cases.

What the studies must establish

Caltech’s use of unpublished records and machine-verified proofs addresses two weaknesses in conventional AI benchmarks: training-data contamination and subjective grading. The medical projects examine whether compact open models can operate under constraints such as institutional data governance, limited connectivity, and available hardware.

Across the clinical work, teams combine a locally hosted language model with specialized components for vision or translation, then retain human review for consequential decisions. This modular design lets institutions choose tools for each data type while keeping sensitive records under local control.

Meta has not disclosed funding by project, and the announcement includes no clinical accuracy results, comparative baselines, or publication schedule. Published studies will need to report error rates, evaluation methods, hardware requirements, and performance against established clinical or scientific workflows. Clinical systems will also require prospective validation, security review, and any applicable regulatory clearance before routine use.