An abdominal CT scan is a thousand-page book in which one sentence matters. A radiologist gets a few minutes to find it, then moves to the next book, and the next, for an entire shift. Alibaba’s research arm just published a machine that reads those books quickly, claims it finds the sentence more often than most humans do, and then did the genuinely unusual part: it gave the machine away.
Damo Academy’s model, released in 2026, identifies nearly 150 abdominal conditions from CT scans, cancers included. It was tested on 40,000 scans and, per Alibaba, outperformed most radiologists. The weights are open. Anyone with a GPU and a hospital badge can start poking at it today.
That last sentence is the story. The accuracy claim is interesting. The license is what actually changes anything.
What the numbers do and don’t tell you
Forty thousand scans is a serious evaluation set by medical imaging standards. “Outperformed most radiologists” is a phrase doing a lot of quiet work, and I’d want the details before I got excited. Most radiologists on which task? Reading under what conditions? With how much clinical context, or none at all?
Human radiologists in a comparison study are usually working blind, with no patient history, no prior imaging, no lab results, no conversation with the referring physician. The model is also working blind, which makes the comparison fair in a narrow sense and misleading in a practical one. In a real reading room, the human has all of that context. The gap on paper tends to shrink once you put the human back in their actual job.
None of that makes the result unimpressive. A single model covering roughly 150 conditions across one anatomical region is a wide net, and wide nets are harder than they look. Narrow detectors trained on one cancer type are relatively well-trodden. Something that reads a whole abdomen and flags anything from the common to the rare is a different engineering problem, and the fact that it holds up across 40,000 scans suggests it isn’t a demo.
Open weights are the real release
Alibaba positioned this as support for clinical workflows and better diagnostic accuracy across varied medical settings, with the open release meant to allow integration into existing radiology systems. Read that as: hospitals that could never afford a licensed commercial detection suite can now run one.
Alibaba and MiniMax have both been pushing open models this year with the stated goal of lowering costs for developers worldwide. In consumer AI, that mostly means cheaper chatbots. In medical imaging, it means a district hospital with one overworked radiologist and no procurement budget gets access to the same tooling as an academic center. That is a meaningfully different kind of cost reduction.
It also means the model is now outside anyone’s control, including Alibaba’s. Open weights can be fine-tuned, forked, quantized, wrapped in a janky web app, and deployed by someone with more enthusiasm than clinical governance. Both the good and the bad version of this future arrive through the same door.
The part nobody puts in the press release
Open weights are not a product. Between a downloadable model and a scan that changes someone’s treatment sits a pile of unglamorous work:
- Regulatory clearance. Open source does not confer approval. Every jurisdiction has its own path, and a GitHub release starts that clock rather than skipping it.
- Integration. Radiology runs on PACS systems, DICOM quirks, and workflows built over decades. “Allows integration” and “integrates” are separated by months of engineering.
- False positives. A model that flags 150 conditions has 150 ways to be wrong. Every spurious flag becomes a follow-up scan, a biopsy, a frightened patient. Sensitivity that looks great on a benchmark can generate real downstream harm at scale.
- Distribution shift. Scanner hardware, imaging protocols, and patient populations vary enormously. Performance on the evaluation set is a starting hypothesis about performance in your hospital, not a promise.
- Liability. When a free model misses a tumor, the question of who answers for it has no settled answer anywhere.
My read
I review AI tools for a living and I spend most of my time deflating claims. This one I’d rather see tested than dismissed. The failure mode of medical AI hype is usually a closed model with a glossy accuracy chart and no way for outsiders to check it. Open weights invert that. Independent researchers can now run this thing against their own data and publish what they find, which is how the claim either holds up or falls apart in public.
Skepticism about the benchmark is warranted. The release model deserves credit. If you work in imaging informatics, the useful response isn’t to argue about whether it beats radiologists. It’s to download it, run it against your own archive, and find out what it does on your scanners with your patients. That answer is now available to you, which it usually isn’t.
đź•’ Published: