NVIDIA releases an open 3D CT reasoning model
NVIDIA has released NV-Reason-CT, an open vision language model that processes CT scans as true 3D volumes and produces radiologist-style reasoning traces, targeting a modality frontier models handle poorly.
NVIDIA has introduced NV-Reason-CT, a vision language model built to read full 3D computed tomography volumes and explain its findings in radiologist-style chain-of-thought, according to a post on the company's developer blog.
The pitch is architectural. Most VLMs flatten a CT study into independent 2D slices, discarding the through-plane continuity that defines masses, effusions, and infiltrates — a problem NVIDIA frames against studies that can run 300–600 axial slices. NV-Reason-CT instead pairs a Qwen3.5-4B language model with a full 3D vision transformer adapted from Primus and initialized with Colipri weights, retrained end to end. Volumes are resampled to 192³ voxels at 2 mm isotropic resolution and cut into non-overlapping 8×8×8 patches, yielding 13,824 vision tokens that are passed to the LLM without merging or downsampling, along with 3D grid coordinates handled via 3D MRoPE.
Capabilities listed include structured report generation against an internally curated ontology of 30 chest and 29 abdominal abnormalities, step-by-step reasoning traces with stated uncertainty, and multiturn follow-up questioning. The work extends the methodology of NV-Reason-CXR, which NVIDIA says was validated in a multireader clinical study accepted at RSNA 2026 showing radiologist time savings without loss of diagnostic accuracy.
NVIDIA is explicit that this is a research and development foundation model intended for post-training, not an autonomous diagnostic system or a cleared clinical product. The available excerpt of the post ends before its benchmark results, so published accuracy numbers are not covered here.