San Francisco-based healthcare AI company Knit Health —founded in 2025 by a team of scientists spun out of the University of California, Berkeley—announced a $35M Series A financing round led by New Enterprise Associates (NEA) , bringing its total capital raised to $46.6M.
The Series A syndicate includes institutional investors Town Hall Ventures and SignalFire alongside returning seed and pre-seed backers Frist Cressey Ventures , Uncork Capital, Moxxie Ventures , Coalition Operators , and The Aligned Fund.
Concurrently with the funding, Knit Health unveiled Med-BEAM, described as the first foundational world model for healthcare, trained on billions of real clinical decisions and hundreds of millions of longitudinal, fully de-identified patient records.
Unlike general-purpose large language models (LLMs) that predict the next word in unstructured clinical text, Med-BEAM learns directly from clinical action sequences—orders placed, referrals sent, diagnoses reached, and downstream outcomes—to construct a living model of each patient’s clinical trajectory and simulate how alternative interventions alter outcomes.
Rather than predicting the next word in a clinical note, Med-BEAM is engineered as a foundational world model trained on billions of real-world healthcare decisions across hundreds of millions of longitudinal, de-identified patient records. By learning directly from empirical sequences of orders, diagnoses, specialist referrals, and downstream outcomes, the model constructs a dynamic, living representation of an individual patient’s clinical state. This allows care teams and software applications to forecast what care a patient is likely to need next and simulate how alternative clinical interventions could alter their health trajectory.
To evaluate its world-model architecture against general-purpose foundation models, Knit Health benchmarked Med-BEAM against a suite of frontier lab LLMs on an unprompted inpatient prediction task—requiring each model to forecast the clinical actions that would occur for hospitalized patients over a 24-hour window without nominating target outcomes in advance.
Source link







