Realtime AI News
InclusionMed launches as an independent benchmark hub for medical AI, backed by 29 institutions
A new platform called InclusionMed has gone live to evaluate medical AI independently, publishing a benchmark atlas and a first index built with researchers from 29 institutions including Stanford, Harvard, Peking Union Medical College and Ant Group. Its first index compares 22 models from 11 vendors across five benchmarks and 1,034 tasks.
Medical AI evaluation has long been fragmented: nearly every narrow task has its own benchmark, yet very few of them answer the question a hospital actually asks — does this model hold up in real clinical work? InclusionMed, newly launched and covered by QbitAI, is an attempt to answer it. The platform is being described as a medical-health equivalent of Artificial Analysis, and it arrives with a benchmark atlas, InclusionMed Atlas, and a first index, InclusionMed Index.
The project was built by researchers from 29 institutions, among them Stanford, Harvard, Peking Union Medical College, Ant Group and InclusionMed.AI. That breadth matters: credibility in clinical evaluation is hard for any single vendor or hospital to manufacture, and a multi-party design forces academic, clinical and industry perspectives to check one another.
Atlas is the more interesting structure. It crosses two dimensions — who the tool serves, and what ability is being tested. The first splits into patients and the public, medical professionals and healthcare administrators; the second into knowledge, reasoning and task execution. The result is a nine-cell grid, and every benchmark is filed into the cell it primarily measures, which makes gaps in coverage visible rather than implicit.
The index aggregates results from the benchmarks admitted into that grid. Its first edition focuses on clinical care and healthcare management, drawing on five benchmarks and 1,034 tasks to compare 22 models from 11 vendors at home and abroad. All results went through independent re-testing, and four public benchmarks were selected alongside the team's own InclusionMed-DxFiles to form the inaugural index.
DxFiles deserves separate mention. It is built from difficult cases published in top medical journals and tests whether a model can combine scattered case clues into a clinical judgement, rather than recite textbook conclusions. That is closer to how doctors actually work, and it is where models tend to reveal weaknesses in reasoning under incomplete information.
The gap this fills is practical. Medical AI procurement has lacked third-party comparison, leaving hospitals unable to judge where a model's competence ends and giving vendors an incentive to publish results selectively. An index that spans multiple roles and capability types, with independent re-testing, at least gives buyers a shared vocabulary.
A benchmark platform is not, however, proof that any model is fit for clinical use. Scores are one slice of capability; deployment still runs through data compliance, physician oversight and liability. The value of InclusionMed is that it moves the conversation from which model is smartest to which task, for whom, and how reliably.
According to QbitAI, the team is now working with medical institutions and research groups on new tasks covering shared patient-physician decisions in complex treatment choices and care scheduling under resource constraints. Whether the benchmark system can keep pace with medical AI itself will decide how much it matters over time.
Why it matters
Hospitals have had no neutral way to compare medical models; a multi-institution index gives procurement a reference point and puts pressure on selective vendor reporting.
Nearby Updates
All11/18, 15:44
California Sues OpenAI, Citing 80-Plus AI Agent Loss-of-Control Incidents
California has sued OpenAI, arguing the company tested its AI models without adequate safeguards and pointing to more than 80 disclosed loss-of-control incidents involving AI agents. Multiple U.S. outlets link the case to an attack involving Hugging Face, making it a direct legal challenge to frontier AI safety practice.
11/18, 15:53
GPT 6.2难产!OpenAI怕了
GPT 6.2难产!OpenAI怕了. OpenAI也终于活成了2022年谷歌的样子
11/18, 16:46
Moore Threads completes Kimi K3.1 Pro training adaptation, holding 400 stable steps on 16 GPUs
Moore Threads has completed a training adaptation for Kimi K3.1 Pro and held 400 consecutive training steps stable on two machines with 16 GPUs, according to a report carried by East Money. The run moves domestic accelerator validation from inference into the harder problem of multi-node training.
11/18, 14:01
Microsoft Foundry Adds Baseten and DeepSeek Inference, Widening Its Open-Model Menu
Microsoft Foundry has added Baseten and DeepSeek AI inference, letting developers reach Baseten-hosted open models and DeepSeek inference services from the same platform. The move continues Foundry's strategy of aggregating multiple model providers under one catalog, billing and governance layer.