NV-Reason-CT in the browser

Runs NV-Reason-CT from NVIDIA, a 3D CT vision-language model, to write structured chest or abdominal CT reports and answer questions about the scan.

Research use only

This is a research demonstration and is not a medical device. It has not been validated or approved for clinical use, and it must not be used for diagnosis, treatment or any other clinical decision. The model can miss findings and can describe findings that are not there. The weights are compressed to 8-bit or 4-bit to fit in the browser. The 8-bit build closely matches the original model; the 4-bit build is smaller but its reports can differ more.

Your images stay on this computer

Files you choose are read and processed inside this browser tab, using your GPU. No image data, questions or answers are uploaded anywhere. The only downloads are this page and the model files (5.0 GB for the 8-bit build or 3.0 GB for the 4-bit build, from Hugging Face), which are fetched once and cached in the browser. After the first visit the page also works with the internet turned off.

What you need

  • Browser: a recent desktop Chrome or Edge with WebGPU. Firefox and Safari are untested. Phones and tablets are not supported.
  • GPU: a desktop or laptop GPU with at least 8 GB of memory for the 8-bit build (12 GB or more recommended), or 6 GB for the 4-bit build. Integrated graphics with shared memory may work but will be slow.
  • Memory: 16 GB RAM minimum. The tab peaks at about 6 GB while the models are loaded.
  • Storage and network: 6 GB of free disk for the browser cache (3.5 GB for the 4-bit build), and an internet connection for the first visit only.
  • Time: on a desktop GPU, loading a scan and writing a report takes about a minute or two; follow-up questions are quicker. There is no CPU mode: without WebGPU the model would take hours per report.

Drop a CT here: a zip of a DICOM study, a DICOM folder, loose DICOM files, or a .nii / .nii.gz volume.

The volume is resampled to 2 mm and a 38 cm cube is cropped around the chest or abdomen, placed using the lungs. From a study, the largest axial CT series whose description matches the chosen region is used; otherwise the largest CT series.

Log