A compact, open-source AI model trained specifically on rare disease cases beat rivals 20 times its size at helping doctors reach the right diagnosis.
Researchers released RaDaR (Rare Disease navigatoR), a 32-billion-parameter reasoning model trained on roughly 49,000 real clinical cases and over 104,000 synthetic ones. In benchmarks and tests across four external validation centers, it outperformed DeepSeek-R1 — a model with 671 billion parameters — on rare disease diagnosis tasks. In a retrospective cohort, RaDaR flagged the correct diagnosis before physicians had formally documented their suspicion in 61 percent of cases, suggesting it could shave about 1.87 months off the typical diagnostic delay.
The randomized trial is the part that matters most: physicians using RaDaR improved their diagnostic accuracy by 21.44 percentage points compared with those who only had internet search. For rare diseases — where the average patient waits years for a correct diagnosis and sees multiple specialists — that gap is clinically significant. The model is open-source and compact enough to deploy without hyperscale infrastructure, which addresses a real barrier previous AI diagnostic tools have run into.
The synthetic-data approach — using phenotype-anchored narratives to fill in where real case data is scarce — is worth watching; if it holds up to scrutiny, it could be a template for building clinical AI in other data-sparse corners of medicine.