Alibaba gives away a CT reader that beat 23 of 26 radiologists
Damo Academy put a Science-validated generalist that reads abdominal CT at 0.913 AUC on Hugging Face, days after it beat 23 of 26 radiologists. The weights are free; the license is non-commercial, and that split decides who actually gets squeezed.
Vincent Jiang · 2 min read
A radiologist you can download
The work hospitals pay for, reading an abdominal CT, has a new public benchmark. Alibaba's Damo Academy released RADAR on 18 September, a vision-language model that reads contrast-enhanced abdominal CT across 18 organs and flags 146 clinical findings, including malignant tumors 12. In a reader study spanning multiple hospitals, its average accuracy exceeded that of 23 of 26 radiologists 17. Working with RADAR as a second reader, those radiologists cut missed diagnoses by about 10% and reading time by more than 30% 12.
Zero labels, 15 million pairs
The asset vendors charge for is the hand-labeled dataset. RADAR skipped it. The model learned from 424,911 existing examinations, 1.5 million image-text pairs and more than 15 million anatomy-specific pairs, with no manual tagging, mined from a decade of reports at a single high-volume center and published in Science 348. Validation held off campus: across eight independent hospitals, AUC ran 0.874 to 0.912, against 0.913 internally 3.
Seventeen findings, cleared and billed
The same month, the regulated frontier moved. The FDA cleared a2z Radiology's Abdo-Triage to flag ten more findings on abdomen-pelvis CT, bringing its total to 17; the Boston startup, founded in 2024, has raised $5M 5. It sells what RADAR lacks: clearance, triage workflow, someone to hold liable 5.
The open model covers 146 findings; the FDA-cleared tool flags 17
Data
| Value | |
|---|---|
| DAMO RADAR (open weights) | 146 |
| Abdo-Triage (2026 clearance) | 17 |
| Abdo-Triage (2025 clearance) | 7 |
The gap in that chart is the market's shape. One system covers 146 findings free for research; the other covers 17 findings and can be billed at the bedside.
Free weights, not a free product
Alibaba's own repository carries the catch. The project ships under CC BY-NC-SA 4.0, and the README states the model is "currently intended for research purposes only," with prospective clinical studies still required before clinical deployment 4. It reads contrast-enhanced abdominal CT and nothing else, and it holds no FDA clearance, so a US hospital running it diagnostically would be operating outside regulatory boundaries 46.
Who deploys it first
Hospitals outside the FDA's reach can download the checkpoint today; US vendors keep their billing codes, for now 456. What the free model already repriced is labor: junior radiologists assisted by RADAR performed at or above the level of seniors working unaided 3. The weights were never the moat. The paperwork was.
How this brief was made
01Gathered & sourced361 channels · 1,226 articles▾
Agents swept 361 channels and ingested 1,226 articles, then de-duplicated and ranked them for signal.
02Verified & cross-validated8 claims · 20 data feeds▾
Every one of 8 load-bearing claims was checked against primary sources, with 20 live data feeds reconciling the figures and charts.
- 1South China Morning Post (via MSN), "Alibaba open-sources medical AI model that can detect cancer and nearly 150 conditions", 18 September 2026
- 2Yicai Global, "Alibaba's DAMO Academy Pushes Beyond Single Disease-Detecting AI With New Diagnostic Model", 18 September 2026
- 3Medical Xpress (via MSN), "AI assistant learns from radiology reports to spot problems in abdominal scans", 22 September 2026
- 4GitHub, alibaba-damo-academy/damo-radar repository (README, license, Science citation), updated 22 September 2026
- 5Radiology Business, "FDA clears radiology AI triage tool to check for 10 additional findings on CT scans", 23 September 2026
- 6Tech Times (via MSN), "Alibaba radiology AI outperforms 23 of 26 radiologists across 146 diseases in Science", 19 September 2026
- 7Milliyet, "Alibaba open-sources Damo Radar to detect cancers and other abdominal conditions", 28 September 2026
- 8Developments Today, "AI Assistant RADAR Learns From Radiology Reports to Read Abdominal CT Scans", 22 September 2026
03Reviewed & edited1 human editor▾
One editor read the draft against the evidence, tuned the framing, and signed off before it shipped.
Become a contributor
Reporting on the business of AI and want it read? We take pitches from outside contributors who bring primary sources and a number worth arguing about.
Deepdive
AI-generated from this story and its cited sources. Not investment advice.
