§ Claim under review · Release
"Alibaba has open-sourced an AI model capable of detecting cancer and nearly 150 other medical conditions from diagnostic data." (Instagram, @aigramapp, 2026-09-19, citing South China Morning Post via news.ycombinator.com)
Verdict
Mostly accurate
Confidence
HighSummary
This one mostly checks out. On 17 September 2026 Alibaba's DAMO Academy published a model called RADAR in the journal Science, and on 18 September it posted the code on GitHub and the model checkpoints on Hugging Face. The paper reports the model identifies 146 findings across 18 abdominal organs from contrast-enhanced CT scans, including liver, pancreas, stomach and colorectal cancers, and that it outperformed most of the 26 radiologists in a comparison study. Two details in the post are off. The cancers are counted inside the 146 findings, so it is not "cancer plus nearly 150 other conditions," and the model works on abdominal CT scans specifically, not on general "diagnostic data." Several outlets also report that while the code is fully open source, the model weights carry a licence that allows research use but bars commercial deployment, which is a real limit on the post's claim that the release democratizes access. The model is a research release and has not been shown to be cleared by any regulator for clinical use.
The readings
key figures from the evidenceRADAR mean AUC on internal cohort, 95% CI 0.911-0.915
total imaging findings across 18 organs, cancers included
Why this verdict
Evidence
Alibaba DAMO Academy, with the First Affiliated Hospital of Zhejiang University School of Medicine and other clinical partners, published a model called RADAR (marketed as DAMO RADAR) in Science on 17-18 September 2026, and released code and model checkpoints publicly on GitHub and Hugging Face on 18 September 2026.
The Science abstract states the model was evaluated across 18 anatomical structures and 146 imaging findings, reaching an AUC of 0.913 (95% CI 0.911 to 0.915) on a real-world internal cohort of 39,160 examinations, and AUCs of 0.874 to 0.912 across eight external centres. For pathology-confirmed evaluation of four cancer types (liver, pancreas, stomach, colorectum), AUCs were 0.891 to 0.984. On acute abdominal conditions excluded from initial training, AUC was 0.904. In a reader study with 26 radiologists from 14 centres, the paper reports the model outperformed most participants. The AAAS release states RADAR's mean AUC of 0.913 compares with 0.776 for existing specialist AI models.
The GitHub README describes RADAR as a generalist vision-language model trained on over 400,000 contrast-enhanced abdominal CT examinations with 15 million anatomy-aware image-text pairs, learning from clinical reports without manual annotation, and states: "This project is released under the Apache License 2.0." Checkpoints and supporting files are hosted on Hugging Face. Multiple tech outlets report that the weights on Hugging Face carry a CC BY-NC-SA 4.0 licence, which permits research use but bars commercial deployment. I was not able to retrieve the Hugging Face licence field directly.
Findings
✓ What's accurate 5
- Alibaba's research arm, DAMO Academy, did publicly release an AI model for medical diagnosis, with code on GitHub under Apache 2.0 and checkpoints on Hugging Face, on 18 September 2026.
- The model detects cancers. Pathology-confirmed evaluation covered liver, pancreas, stomach and colorectal cancers with AUCs of 0.891 to 0.984.
- The figure of roughly 150 conditions traces to a real number: 146 imaging findings across 18 anatomical structures.
- The underlying work is peer-reviewed and appeared in Science, not a preprint or a blog post.
- SCMP did report the story, so the post's cited chain is genuine.
≈ What's misleading 4
- Exaggeration: the post says "cancer and nearly 150 OTHER medical conditions." The evidence says 146 findings in total, with cancers counted among them. The added word "other" converts a total into a total-plus-cancer and inflates the scope by the most attention-grabbing category.
- Omitted qualifier: the post says "from diagnostic data." The model reads contrast-enhanced abdominal CT scans only, covering 18 abdominal organs. A reader is left to imagine a general-purpose diagnostician working from any medical data, which the evidence does not support. This is the single largest gap between the post and the paper.
- Definitional dispute over "open-sourced": the code carries Apache 2.0, an OSI-approved licence, but the weights are reported by several outlets to carry CC BY-NC-SA 4.0, which forbids commercial use. If that holds, the release is open-weights-for-research rather than open source in the full sense, and the post's framing that it will "democratize access" to medical AI tools overstates what a hospital or startup could legally deploy. The remaining disagreement here is definitional, but the underlying licence facts are what matter.
- Demo to product conflation: "capable of detecting" reads as a deployable clinical capability. The evidence is retrospective cohort evaluation plus a reader study. Regulatory clearance and prospective outcome data are not established by this release.
? What's uncertain 4
- The exact licence on the Hugging Face weights. Three secondary outlets state CC BY-NC-SA 4.0, and one notes the GitHub LICENSE file says Apache 2.0 while a badge in the same README points to CC BY-NC-SA 4.0. I could not retrieve the Hugging Face licence field itself, so this is reported, not verified at the artifact.
- Whether the release includes full training weights for all variants (RADAR and RADAR+) or a subset. The repository references both, and I did not enumerate the checkpoint files.
- Independent reproduction of the reported performance. None exists yet; the release is two days old as of 2026-09-20.
- Regulatory status. No clearance for clinical use was found for this specific model. A separate earlier DAMO pancreatic-cancer model is referenced elsewhere as having US FDA Breakthrough Device designation, which is a different artifact and does not transfer to RADAR.
Sources
7 of 7 linked to records"An expert-level generalist AI for abdominal CT diagnosis," Science, vol. 393, issue 6817, eaec6129, published 17-18 Sep 2026
GitHub repository alibaba-damo-academy/damo-radar, code and README including licence statement, updated 18 Sep 2026
Zenodo archival record of the RADAR source code, pointing to the Hugging Face page huggingface.co/radar-generalist
Hugging Face organisation page "radar-generalist" hosting checkpoints
AAAS/EurekAlert press release, "Introducing RADAR, a generalist AI tool for abdominal CT diagnosis," 17 Sep 2026
South China Morning Post, "Alibaba open-sources medical AI model that can detect cancer and nearly 150 conditions"
Tech Times (19 Sep 2026), Intelligent Living, and antihype.com.br, each reporting the weights licence as CC BY-NC-SA 4.0 while code is Apache 2.0