A model that beat most radiologists across 40,000 CT scans is now free to download. Almost nobody reading this is allowed to point it at a real patient.
Both things are true, and the gap between them is the whole story. Alibaba’s research arm, Damo Academy, open-sourced an AI model in 2026 that identifies nearly 150 abdominal conditions from CT scans, cancers included. It was tested on 40,000 scans and outperformed most radiologists. The weights are out. The paperwork is not.
I review tools for a living, which mostly means asking one unglamorous question over and over: what happens when a normal team tries to actually use this? So that’s the angle here, not the medical breakthrough framing you’ll find everywhere else this week.
What was actually released
The verified specifics are narrow but meaningful. Damo Academy built the model to support clinical workflows and improve diagnostic accuracy across different medical settings, and open-sourced it so it can be integrated into existing radiology systems. That last detail matters more than the accuracy number. A closed API that scores well is a vendor demo. Weights you can run inside your own infrastructure, next to your own PACS, is a different category of thing entirely.
It also fits a pattern. The same month, Alibaba and MiniMax were both pushing new open models out the door with the stated goal of lowering costs for developers worldwide. Medical imaging showing up in that release cadence is the part worth watching. Diagnostic AI has historically been the most locked-down, license-gated corner of the field. Now a serious one is sitting in a public repo.
The part the headlines skip
“Outperformed most radiologists” on a benchmark set is not the same as outperforming a radiologist in your hospital, on your scanner, with your patient population. Every reviewer who has watched a model crater the moment it left its training distribution knows this reflex. The 40,000-scan evaluation is a strong signal. It is not a deployment guarantee, and I have no data on how the model behaves on equipment or demographics outside that test.
Then there’s the boring wall that stops almost every clinical AI project:
- Regulatory clearance. Open weights do not come with approval from any medical device regulator. Whoever deploys this owns that process.
- Liability. If the model misses something, the license does not absorb the consequences. The clinician and the institution do.
- Validation cost. Local validation on your own scans is the actual price of admission, and it dwarfs the cost of the model itself, which is now zero.
- Integration reality. “Integrates into existing radiology systems” is an architectural claim. Anyone who has touched hospital IT knows how much work hides in that sentence.
None of that makes the release less interesting. It just means the free part is the cheapest part.
Who this is genuinely useful for
Three groups, in my read.
Researchers get the biggest immediate win. A high-performing baseline for abdominal imaging that anyone can inspect, fine-tune, and criticize is more useful to the field than a closed model with a better score. Reproducibility is not a side benefit here, it’s the point.
Health systems in cost-constrained regions get an option they did not have before. Damo Academy explicitly aimed at working across diverse medical settings, and the economics of open weights favor places where per-scan licensing from a Western vendor was never realistic. That’s where I’d expect the first serious deployments.
Commercial diagnostic AI vendors get a problem. When a free model performs at this level, selling the model stops working. Selling validation, monitoring, regulatory support, and integration starts working. That shift has already happened in other software categories, and it tends to be good for buyers.
My honest verdict
This is the most consequential open release I’ve looked at this year, and also the one least likely to show up in anybody’s stack next quarter. Both of those are correct at the same time.
What I like: the decision to open-source instead of licensing. Diagnostic models that nobody outside the vendor can audit have been a quiet problem in this field, and Alibaba just made auditing possible for one of the strongest ones.
What I’d want before recommending it to anyone: independent evaluation on non-Chinese patient cohorts, failure mode documentation, and some honest reporting on how it degrades on lower-quality scans. I don’t have any of that yet, and I’m not going to pretend otherwise.
If you build medical imaging tools, download it this week and start testing. If you run a clinic, the model is the easy part. Budget for everything after it.
🕒 Published: