Intelligence upward means constantly pushing the limits of AI capabilities; intelligence for good means ensuring that every technological leap addresses real clinical needs. In July 2026, Unisound released the U2-Med three‑domain medical foundation model, extending large‑model capabilities across the full spectrum of healthcare, medical insurance, and pharmaceuticals. Just one month later, Unisound’s medical AI capabilities have evolved once again—with the official launch of its proprietary medical imaging multimodal foundation model, U2‑RadiMed.

This rapid iteration is fueled both by deep technical reserves and by the opportune industry window for medical imaging AI: the high‑growth cloud imaging market, accelerating policy implementation for mutual recognition of medical imaging across insurance schemes, and the growing clinical urgency for AI‑assisted reading. This time, we are tackling the more complex and specialized challenges of medical imaging head‑on.

As the first domestic imaging foundation model to deeply integrate medical image understanding, report generation, clinical reasoning and decision‑making, and visual question answering in a full‑chain solution, U2‑RadiMed supports multimodal mixed input of single or multiple images and text. It completes a closed‑loop workflow of “view images → retrieve knowledge → reason → conclude,” providing clinicians with comprehensive AI assistance—from lesion screening and precise diagnosis to personalized treatment decisions—thus propelling imaging diagnosis to a higher level of intelligence.

U2‑RadiMed fully covers mainstream medical imaging modalities including CXR (chest X‑ray), CT, and MRI, and establishes a six‑pillar core capability framework: report generation, lesion detection and localization, clinical reasoning, clinical decision‑making, general medical VQA, and advanced medical reasoning. It is a “radiology AI expert” equipped with professional diagnostic thinking.

I. Deepening Iterative Upgrades in Medical AI – The U2 Series Complements Core Imaging Cognition Capabilities

U2‑RadiMed is the outcome of Unisound’s large‑model capabilities continually “growing deeper” in the medical field.

Leveraging its continuously evolving technical system, Unisound has progressively extended general intelligence into specialized medical scenarioses: from the general‑purpose foundation model U2, to U2‑Med covering the “three‑medical” domains (healthcare, insurance, pharmaceuticals), and now to U2‑RadiMed for in‑depth medical image understanding—each step brings large‑model capabilities into more complex and more professional real‑world clinical settings.

If we liken this evolution to training an “AI physician”:

• U2 taught it to think,

• U2‑Med taught it to understand medicine,

• and U2‑RadiMed gives it a pair of “eyes that can truly read medical images.”

Built upon the professional medical capability foundation of U2‑Med, U2‑RadiMed further integrates visual understanding, linking lesion features in images with medical knowledge, clinical information, and diagnostic logic.

Thus, when faced with a CT scan or a set of MRI images, it no longer merely answers “what is in the image,” but goes further to analyze around clinical questions:

• Where is the lesion? What features does it present?

• What might it imply? Which diseases need to be differentiated or ruled out?

• What should be the next step in evalsuation?

From “seeing images” to “understanding images,” and then to reasoning and judging with medical knowledge, U2‑RadiMed bridges the gap between “medical vision” and “medical cognition,” continuously evolving Unisound’s medical foundation model toward more professional and deeper clinical core scenarioses.

II. Continuously Advancing Technology – Multiple Core Capabilities Lead in evalsuations

A top‑tier medical imaging foundation model must achieve not only precise visual recognition but also accurate localization, clear Q&A, and deep reasoning. In multiple horizontal evalsuations, U2‑RadiMed has demonstrated comprehensively leading hard‑core strength.

In the general comprehensive evalsuation, U2‑RadiMed ranked first with a composite score of 55.5, surpassing Gemini 3.1 (47.9), GPT‑5.4 (47.5), as well as domestic and international models such as Hulu‑Med‑27B and Qwen3.6‑27B, showcasing a more balanced multimodal medical imaging capability.

In the six core capability sub‑evalsuations, U2‑RadiMed achieved top scores in four capabilities: report generation, lesion detection and localization, general medical VQA, and advanced medical reasoning. Its advantage is most pronounced in lesion detection and localization, with a score of 37.6—significantly outperforming peer models—establishing a clear competitive edge in the critical aspects of lesion abnormality screening and precise localization.

Meanwhile, the model scored 74.2 in general medical VQA and 46.8 in advanced medical reasoning, both ranking first among evalsuated models. Clinical decision‑making, clinical reasoning, and report generation together form a complete closed‑loop capability from lesion identification, image interpretation, and intelligent Q&A to clinical reasoning and diagnostic decision‑making.

In the previously released MedBench 5.0 full‑modal foundation model evalsuation, U2‑RadiMed also ranked first with a composite score of 57.2, further confirming its top‑tier comprehensive multimodal medical imaging capability.

III. Breaking Through the Limitations of Single‑Frame Imaging – Understanding Temporal Disease Courses and Clinical Diagnostic Logic

Real‑world clinical practice never analyzes a single image in isolation.

Physicians need to combine imaging changes over different time points, clinical information, and medical knowledge to make integrated judgments about disease progression, treatment efficacy, and subsequent plans.

U2‑RadiMed thus goes beyond the traditional limitation of single‑image analysis. It can understand the relational logic between multi‑temporal images, supporting diverse clinical scenarioses such as lesion detection and localization, intelligent image Q&A, multi‑temporal image comparison, dynamic disease tracking, differential diagnosis, and assisted report generation—closely mirroring the real diagnostic thinking and workflow of clinicians.

Case 1: Temporal imaging links a 7‑year disease course, enabling full‑cycle tumor analysis

For MRI images of a patient with low‑grade glioma at three stages—initial diagnosis, pre‑surgery, and post‑surgery—U2‑RadiMed abandons isolated single‑image interpretation. Instead, it links temporal images to reconstruct the complete disease course. The model accurately identifies the 25 mm left‑frontal lesion at initial diagnosis 7 years ago, the progression to 45 mm pre‑surgery, and determines that there is no significant residual mass post‑surgery, confirming total resection. Moreover, it can precisely distinguish postoperative hyperintense edema zones as normal postsurgical changes and, incorporating the pathological features of low‑grade glioma, provide professional postoperative follow‑up recommendations. This capability transforms static images into a dynamic disease narrative, enabling full‑cycle AI assistance covering disease progression, treatment response, and postoperative management.

Case 2: Full‑chain reasoning and traceability for interpretable precise differential diagnosis

In a clinical case of a 70‑year‑old male with bilateral proptosis, orbital contrast‑enhanced CT shows prominent retro‑orbital fat and thickened extraocular muscles, with characteristic sparing of the tendinous insertions while muscle bellies are enlarged. For this case, U2‑RadiMed does not limit itself to a single diagnosis; instead, it performs differential diagnosis based on complete clinical logic.

First, the model combines the core imaging features to rank thyroid‑associated ophthalmopathy (TAO) as the most likely diagnosis, while systematically ruling out possible conditions such as idiopathic orbital inflammation, orbital lymphoma, and orbital vascular lesions—listing supportive and exclusionary evidence for each. Based on this, it further recommends standardized next‑step diagnostic and treatment measures, including thyroid function and antibody tests, orbital MRI re‑evalsuation, and, if necessary, whole‑body imaging assessment and tissue biopsy. The entire process forms a complete reasoning chain of “imaging feature recognition → medical knowledge retrievals → multidimensional differential diagnosis → risk probability assessment → subsequent diagnostic and therapeutic guidance,” offering strong clinical interpretability and practicality.

IV. Upward Intelligence, Always for Good – Making Medical Imaging AI Accessible to Clinical Practice

Technology continues to advance upward, but its value always aims for good. U2‑RadiMed not only pursues breakthroughs in medical imaging AI capabilities but is also committed to bringing professional capabilities into clinical practice and serving physicians.

In addition, U2‑RadiMed will soon be available on the Unisound MaaS platform (Token Hub), which provides OpenAI‑compatible interfaces, supports multi‑image input and a 40K context window, enabling out‑of‑the‑box usage of the model’s capabilities. Experience link: http://maas.hongre128.com/

From giving AI a pair of “eyes that understand images” to bringing professional imaging capabilities to more hospitals, physicians, and clinical scenarioses, U2‑RadiMed is using “upward” technology to practice “for‑good” values—driving medical imaging AI from technical breakthroughs to true clinical productivity.