RADAR is a generalist vision-language model for diagnosis from contrast-enhanced abdominal computed tomography, developed at Alibaba's Damo Academy with collaborators at Chinese hospitals and universities. It was published in Science on 17 September 2026, with code, model checkpoints and inference and training documentation released publicly (Source: science.org; Source: github.com). Press coverage of the release generally refers to the model as Damo Radar, after its repository name; the authors call it RADAR, for Rapid Abdominal Diagnosis with AI and Radiology (Source: science.org).
The model addresses a modality the paper's editorial summary describes as "one of the more difficult to evaluate because of the number of anatomical structures in the abdomen and the fact that clinically important findings can present as subtle abnormalities that are easily missed" (Source: science.org).
Training and architecture
RADAR is a vision-language model trained on more than 400,000 contrast-enhanced abdominal CT examinations paired with 15 million anatomy-wise image-text pairs, "learning directly from clinical reports without manual annotation" (Source: science.org). The authors position this against prior supervised approaches, which the editorial summary characterises as relying on "supervised learning and human annotations, making them very labor intensive to train." The paper's structured abstract names the training dataset RAD-CT, "consisting of 424,911 examinations and 15 million anatomy-wise pairs," and describes the method as a fine-grained vision-language framework that decomposes CT volumes into anatomical units and aligns them with the corresponding report text through adaptive contrastive learning (Source: science.org).
Parameter count is not disclosed in the abstract or repository. The implementation builds on four open-source projects — LAVIS, nnU-Net, MONAI and 3D-ResNets-PyTorch — whose licences are retained separately in the repository. Two model variants are distributed, RADAR and RADAR+; the repository documents training from scratch, fine-tuning on MERLIN data, and inference against both an internal RAD-CT demo and the external MERLIN test set (Source: github.com).
Capabilities and evaluation
The authors report that across internal and external evaluations spanning multiple centres and clinical scenarios, RADAR "achieved high diagnostic performance and robust generalization for 18 anatomical structures and 146 imaging findings," and that in a reader study "RADAR assistance increased the diagnostic sensitivity of 26 radiologists by ~10%" (Source: science.org). The paper's stated conclusion is that "generalist AI can match human experts in general and complicated radiology tasks." The editorial summary adds that the model performed well "even for abnormalities it had not encountered during training."
The paper's structured abstract gives the underlying figures. On a real-world internal cohort of 39,160 examinations, RADAR achieved an area under the receiver operating characteristic curve (AUC) of 0.913 (95% confidence interval 0.911 to 0.915) for the diagnosis of the 146 findings, and AUCs of 0.874 to 0.912 across eight external centres. In pathology-confirmed evaluation of four cancer types (liver, pancreas, stomach and colorectum) AUCs reached 0.891 to 0.984. In a separate cross-population cohort the model reached an AUC of 0.883 without fine-tuning, and on acute abdominal conditions, which had been excluded from initial training, it reached 0.904. In a reader study with 26 radiologists from 14 centres, "RADAR outperformed most participants and increased sensitivity by ~10% in collaborative use." The model also generates attention maps that highlight the image regions behind a prediction, which the authors present as an aid to interpretation (Source: science.org).
Coverage of the release carried figures that the structured abstract supports only in part. The institute reported an average area under the curve of 0.913 across 146 clinical findings in nearly 40,000 real-world examinations, that the model outperformed 23 of the 26 radiologists in the reader study on average, and that radiologists using it reduced missed diagnoses by 10% and cut diagnosis time by more than 30%; the team also described the system as "the world's first expert-level generalist medical imaging model" (Source: aa.com.tr). The AUC matches the structured abstract's 0.913 on 39,160 internal examinations. The structured abstract states the reader-study comparison as outperforming "most participants" without giving a count, and neither the 23-of-26 figure nor any reading-time result appears in the publisher's abstracts. Both abstracts state the assistance effect as a ~10% increase in sensitivity rather than a reduction in missed diagnoses, and the unstructured abstract frames its conclusion as parity with human experts (Source: science.org). The South China Morning Post reported that the model "outperformed most radiologists," a comparison its own account placed in the article summary rather than the body (Source: scmp.com); the structured abstract's reader-study result supports that comparison. The figure of "close to 150 conditions" common in coverage is a rounding of the paper's 146 imaging findings.
No safety case, model card, or evaluation of misuse potential accompanies the release.
Availability and licensing
Pre-trained checkpoints and supporting files are distributed through Hugging Face under the radar-generalist organisation, with helper scripts in the repository for fetching checkpoints and processed masks; the code is additionally archived on Zenodo (record 21271172). The repository states a minimum of Python 3.10 (Source: github.com).
The licence terms differ by artefact and host. The repository's licence badge links to Creative Commons Attribution-NonCommercial-ShareAlike 4.0, while its License section and LICENSE file state the Apache License 2.0; the two differ materially, the first barring commercial use and the second permitting it (Source: github.com). The model checkpoint repository on Hugging Face (radar-generalist/RADAR, created 22 July 2026) and its auxiliary-data repository are both tagged CC BY-NC-SA 4.0 (Source: huggingface.co; Source: huggingface.co). The two Zenodo archives also differ. Record 21271172, which the repository README links and which was published on 9 July 2026, carries CC BY 4.0 (Source: zenodo.org). Record 21504519, labelled version v3 and published on 23 July 2026, carries CC BY-NC-SA 4.0 (Source: zenodo.org). The paper cites the v3 record as the deposit of "the model's checkpoint and all code supporting RADAR's training and inference," describing it as made "for unrestricted public access" without naming a licence (Source: science.org). On the hosts' own metadata, the weights are distributed under a non-commercial licence and the GitHub code under a licence that permits commercial use. Coverage describing the release as open-sourced does not distinguish between these terms.
The repository history predates the announcement: its earliest committed directories date from early July 2026 and it carried 45 commits by 21 September 2026, with the 18 September commit updating the README's citation and links. The September dates reported in coverage therefore mark the Science publication and public announcement rather than the first code release (Source: github.com).
Authorship
The paper lists 40 authors led by Qi Zhang, with Qi Zhang, Jianpeng Zhang, Ling Zhang, Wenbo Xiao and Tingbo Liang as corresponding authors, spanning Alibaba's Damo Academy and Chinese clinical and academic institutions (Source: science.org). It appeared in Science volume 393, issue 6817, article eaec6129, under DOI 10.1126/science.aec6129. The first author's listed affiliations include the Department of Hepatobiliary and Pancreatic Surgery at the First Affiliated Hospital, Zhejiang University School of Medicine, in Hangzhou. The manuscript was received on 29 September 2025, resubmitted on 12 February 2026 and accepted on 23 July 2026 (Source: science.org).
The paper lists funding from Chinese national and provincial research programmes, including the Noncommunicable Chronic Diseases-National Science and Technology Major Project, the Major Research Plan and the Science Center Program of the National Natural Science Foundation of China, and several Zhejiang provincial programmes; the first author also acknowledges the K. C. Wong Education Foundation. The paper's competing-interests statement discloses that six authors have filed a related Chinese patent application (CN121526964A), that eight authors are employees of Alibaba DAMO Academy, and that six hold Alibaba stock "as part of the standard compensation package." Patient imaging data are available on request after institutional review board approval and "in compliance with relevant national regulations" (Source: science.org).
Provenance note: the Science paper is a foundational source that has not yet been ingested; its full text is behind the publisher's paywall and was not retrievable on 21 September 2026. Citations to it here are inline
(Source: …)to the publisher's landing page, which carries the abstract, structured abstract and article notes, and should upgrade to the primary text onceneeds-review/INGEST-radar-science-abdominal-ct-2026.mdis resolved.
Relationships
- deploys-in: Healthcare — AI Deployment — diagnostic imaging is the deployment sector
- related: Alibaba / Qwen Team — Damo Academy is Alibaba's research institute
- related: AlphaFold (DeepMind) — comparable specialist scientific model released with open artifacts
- related: Open-Weight Frontier Models — the release's licensing ambiguity bears on what "open" means here
- related: AI Deskilling — reader-study designs of this kind are the evidence base the deskilling literature contests