GenMD

Work with us

The healthcare data access layer for frontier AI

GenMD connects health systems, researchers, and AI labs to privacy-preserving clinical data.

Work with us

Explore benchmarks

01 · Where models break

Frontier models pass medical exams, then struggle with real charts.

Exam-style benchmarks suggest models are ready for clinical work. Real patient records show where they aren't. We find the failures that only appear in real clinical data, and build the data to fix them.

The longitudinal thread

Models read a single note well but lose track across years of encounters, medication changes, and lab trends, which is where most diagnoses are made.

Real-world mess

Copy-forward notes, local abbreviations, and problem lists that contradict each other are routine in real charts and rare in benchmarks.

Across modalities

A radiology finding changes what a lab trend means. Models rarely connect evidence across reports, labs, codes, and notes.

Knowing what it doesn't know

Models answer confidently where a physician would order another test, defer, or escalate. In clinical use, knowing when not to answer matters as much as accuracy.

02 · What's missing

Most benchmarks test a single paragraph. Clinical decisions draw on the whole record.

Most clinical benchmarks give a model a vignette or a single note and score it against an answer key. The failures that matter tend to sit in what those tests leave out.

Today's benchmarks

GenMD Clinical Index

Context

A vignette or a single note.

The full longitudinal record: years of encounters, labs, imaging, and medications.

Source

Exam-style questions, many already published online.

Real, de-identified cases from our data partners, held out from training.

Grading

Multiple-choice keys or preference votes.

Rubrics written and graded by practicing physicians.

Uncertainty

Always rewards giving an answer.

Credits deferring, ordering a test, or escalating when that's the right call.

Failures

One headline accuracy score.

Where and how the model fails, by task and specialty.

Over time

A static snapshot.

Re-tested as records, guidelines, and your model change.

03 · The benchmark

GenMD Clinical Index

Benchmarks that test whether frontier models can handle real clinical work. Built from real cases, written and graded by practicing physicians.

Radiology

Reasoning questions built from real radiology reports, where a model must read the findings and apply one step of clinical knowledge.

Built from

GenMD Nexus Data

Verified by clinicians

Nexus_Benchmark

67 tasks

Single-best-answer questions written from de-identified radiology reports and verified by clinicians. Each task passed validity checks (the key holds against the full report, the answer cannot be guessed without the report) and is graded on the model's chosen option.

Gemini 3 FlashAPI

66.7%±11.4%

Qwen2.5-32B

66.2%±10.7%

gpt-oss-120Bmedium

64.2%±9.9%

MedGemma-27Bq8

61.2%±11.4%

Baichuan-M2-32Bq4

60.7%±9.5%

Nexus_Benchmark · 67

Radiology

Reasoning questions built from real radiology reports, where a model must read the findings and apply one step of clinical knowledge.

Built from

GenMD Nexus Data

Verified by clinicians

Extended

98 tasks

The full suite: the 67 Nexus_Benchmark tasks plus 31 harder tasks, 98 in total, all verified by clinicians and checked the same way. The harder tasks are the ones the reference model never answered correctly.

Gemini 3 FlashAPI

65.6%±9.4%

gpt-oss-120Bmedium

60.9%±8.3%

Qwen2.5-32B

58.8%±9.6%

Baichuan-M2-32Bq4

54.4%±8.4%

MedGemma-27Bq8

51.4%±9.4%

Extended · 98

04 · How we work

Find where models fail, then fix it.

We run a repeating cycle with practicing physicians, using de-identified data from our data partners.

01 · Measure

Find where your model fails

Run your model on physician-written tasks built from real, de-identified cases. Clinicians define what counts as a pass.

02 · Build

Build data for the failures

Physicians author tasks, rubrics, and reasoning trajectories aimed at what broke, drawn from longitudinal, multimodal records.

03 · Re-test

Confirm the improvement

Re-evaluate on held-out, uncontaminated sets. Disagreements between the model and physicians refine the rubric for the next round.

↻ Repeat

Each round raises the bar for the benchmark and improves the training data.

05 · The data

Clinical data for frontier model development.

Longitudinal, multimodal clinical data, curated for pre-training, post-training, and evaluation.

Pre-training

Large-scale, longitudinal patient journeys.

Post-training

Physician-annotated tasks and trajectories.

Evaluation

Uncontaminated clinical sets.

Sourced directly from data partners.

Every dataset comes through a direct partnership with a health system or data provider, is de-identified with validated methods, and stays under the originating institution's governance. Provenance is documented for every dataset.

Become a data partner

Build frontier AI on real-world clinical data.

Work with us

GenMD

Privacy-preserving clinical data for research and frontier AI.

SOLUTIONS

COMPANY

CONNECT