Tasks and inputs
NV-Reason-CT is listed for Image-text-to-text, Visual question answering. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
NV-Reason-CT is NVIDIA's open 4.7B 3D vision-language model for chest and abdominal CT report drafting, Q&A and reasoning, for research use only.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Open weights; no per-token price is published. Your cost is the GPU time you run it on.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
A CT study is a volume, not a picture. Most vision-language models still slice it into 2D images or squash the 3D features before the language model sees them. NV-Reason-CT keeps the volume intact. Its native 3D encoder turns a 384 x 384 x 384 mm region into a 24 x 24 x 24 grid, 13,824 visual tokens, and hands every one of them with its 3D position to the language model.
That is the design bet. Whether it pays off for your scans is a separate question.
| Item | Value |
|---|---|
| Repository | nvidia/NV-Reason-CT |
| Vision encoder | Primus 3D ViT, initialised from COLIPRI |
| Language model | Qwen3.5-4B, with 3D MRoPE positions |
| Parameters | 4.69B total; 4.35B on the active CT path |
| Input | NIfTI volumes (.nii, .nii.gz) plus a text prompt |
| Preprocessing | Auto-crop to chest or abdomen, resample to 2 mm isotropic |
| Licence | OpenMDW-1.1 |
Training combined SFT and GRPO on roughly 550,000 structured QA pairs drawn from 70,111 CT volumes. The data mixes standardised reports, abnormality questions, multi-turn exchanges and reasoning written by radiologists while they read scans aloud. GRPO used verifiable rewards over chest and abdominal abnormality sets.
This is a research and teaching model. It must not be used for diagnosis or treatment decisions, and every output needs review by a qualified clinician. The reasoning text it writes is reviewable output, not a window into its internal computation.
The anatomy cropper relies on lung Hounsfield units and a morphology heuristic. Unusual scan geometries can defeat it. If that happens, crop manually and set anatomy_region=None.
Use de-identified studies your department has already reported. Ask for a draft report on each, then have a radiologist score omissions and false findings separately, because those errors carry different costs. Check the saved crop for a handful of cases to confirm the right anatomy went in. The repository carries no public release date; it appeared on 8 September 2026.
For general image-text work, the image-text-to-text category lists alternatives, and NVIDIA's other releases are on its provider page.
Compare the model on the work you need to complete.
NV-Reason-CT is listed for Image-text-to-text, Visual question answering. It accepts image, text and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under OpenMDW-1.1. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
sensetime
SenseNova-Loop-9B is SenseTime's 9B looped multimodal model: 32 layers run as 56, 512K context, ~15 GiB at 256K in 4-bit, open weights under Apache 2.0.
Tencent
Tencent Hy4 Flash is the low-cost Hy4 tier with open weights: a 220B MoE (~11.3B active) for agents and documents, with 1M context and image input.
shanghai-ai-lab
Sage S2.1 Flash is Shanghai AI Lab's open scientific multimodal agent model, reading molecules, crystals and proteins alongside text and images.
Meituan
LongCat-2.5 is Meituan's open-weight MIT-licensed multimodal MoE for long-horizon agents: screenshots, charts, video and multi-step coding.
sensetime
SenseNova-Loop-9B-Preview is SenseTime's looped 9B text-and-image model for long-context agentic work, running shared layers twice for extra depth.
Liquid AI
LFM3-VL-2B is Liquid AI's open 2B vision-language model for images and video on-device, strong on documents, charts and GUI screenshots.