News

Updates and announcements from NVIDIA MedTech Open Models

New Model

NV-Reason-CT: reasoning on 3D chest and abdominal CT

NV-Reason-CT is a generative vision-language model for native 3D chest and abdominal CT. It pairs a Qwen3.5-4B language model with a 3D Vision Transformer (Primus, initialized from COLIPRI) and reads NIfTI volumes directly.

  • Structured reports — abnormality classification and structured report generation for chest or abdominal CT
  • Reasoning and VQA — step-by-step reasoning, focused questions, and multi-turn follow-ups on the same volume
  • Training recipes — SFT and GRPO examples with region-aware verifiable rewards
  • Agent skill — npx skills add NVIDIA-Medtech/medical-AI-skills --skill nv-reason-ct runs inference from a plain-language prompt

Weights are on HuggingFace, code on GitHub, and the method in the paper. Try it in the web demo, or see example agent prompts on the model page.

Research and development use only — not a medical device, not for diagnosis or clinical decision-making.

Agent Skills

Run the models from your coding agent

We've published Medical AI Skills — agent-callable skills that let a coding agent (Claude Code, Cursor, or Codex) run these models correctly on the first try, from a plain-language prompt. Twelve skills now cover the imaging models:

  • Run skills — segmentation (NV-Segment-CT, NV-Segment-CTMR), synthetic generation (NV-Generate-CT, NV-Generate-MR, NV-Generate-MR-Brain), and chest-X-ray reasoning (NV-Reason-CXR)
  • Finetune skills — continual-learning and finetuning wrappers for NV-Segment-CT, NV-Generate-CT, and NV-Generate-MR-Brain
  • DICOM utilities — preflight, metadata extraction, and DICOM→NIfTI conversion

Install one with npx skills add NVIDIA-Medtech/medical-AI-skills, then ask your agent in plain language — each model page lists example prompts.

Every skill ships with its benchmark report, eval prompts, tests, and output validators, linked from each model page so you can review the evidence yourself — engineering reproducibility, not a clinical claim.

Skills follow the open Agent Skills spec; browse them on GitHub. Engineering and research use only — not a medical device.

GTC 2026

Physical AI and Expanded Imaging Models at GTC

We are releasing 6 new models at GTC 2026, expanding MedTech Open Models from 5 models to 11 models across medical imaging and physical AI.

Physical AI (new category)

  • GR00T-H — Vision-language-action model for surgical robot control, post-trained on the Open-H embodiment dataset
  • Cosmos-H-Surgical — Surgical world model for video prediction and sim-to-real photorealistic video transfer
  • Cosmos-H-Surgical-Simulator — Action-conditioned simulation for in-silico policy evaluation

New Imaging Models

  • NV-Raw2Insights-MRI — Accelerated MRI reconstruction from undersampled k-space (8x–24x), trained on cardiac MRI
  • NV-Raw2Insights-US — Speed-of-sound estimation from raw ultrasound channel data for sharper, tissue-corrected beamforming
  • NV-Generate-MR-Brain — Multi-contrast brain MRI synthesis (T1, T2, FLAIR, SWI)

All models are available on the MedTech Open Models HuggingFace collection with code and documentation on GitHub.

Open Models Hub Released

We are pleased to announce the NVIDIA MedTech Open Models Hub: a collection of open foundation models for medical imaging that segment, reason, and generate across clinical data. Designed to accelerate research and development in medical AI.

Initial Models

  • NV-Segment — 3D CT and MR segmentation across 345+ anatomical regions with automatic detection and interactive point-click refinement
  • NV-Generate — High-resolution synthetic 3D CT and MR volume generation with anatomical annotations for privacy-preserving data augmentation
  • NV-Segment-CTMR — 345+ class automatic segmentation across CT body, MRI body, and MRI brain
  • NV-Generate-MR — Synthetic 3D MRI generation by body region, with optional masks for up to 127 anatomical classes
  • NV-Reason — Vision-language model with chain-of-thought reasoning for chest X-ray interpretation, aligned with radiologist analysis workflows

Model weights are distributed through the MedTech Open Models collection on HuggingFace. The stack is built with MONAI, the NVIDIA co-founded open-source toolkit for medical AI.