Alibaba Open-Sources DAMO RADAR, a Generalist CT Model That Beat 23 of 26 Radiologists
South China Morning Post (corroborated by Yicai Global and Sarawak Tribune) / Hacker News (128pts)·high signal
Alibaba's DAMO Academy published RADAR to GitHub under Apache 2.0 on September 18, one day after the underlying Science paper, with checkpoints on Hugging Face and an archived copy on Zenodo. The vision-language model reads contrast-enhanced CT across 18 abdominal organs and scored an average AUC of 0.913 on 146 clinical findings across roughly 40,000 real examinations; in a 26-radiologist study it outperformed 23 of them on average. Training used more than 420,000 abdominal CT examinations and over 15 million anatomy-focused image-text pairs. Weights plus training framework under a permissive license is the unusual part, not the accuracy claim.