C-RADIO
NVIDIAC-RADIO is NVIDIA Research's commercially licensed line of RADIO vision backbones, which distill several teacher models into one Vision Transformer producing both image-level and dense features. C-RADIOv4 learns from SigLIP2, DINOv3 and SAM3, and adaptors reproduce each teacher's features, including text-aligned embeddings for zero-shot classification. The models load through PyTorch Hub or Hugging Face Transformers.
Openness
3 medium confidence- weights
- open (C-RADIOv4 SO400M and H and earlier C-RADIO checkpoints on the Hub, ungated)
- data
- closed (NV-CC-Img-Text-Dataset, about 700 million images, not released)
- code
- partial (inference, adaptor and evaluation code; no distillation training code)
- license
- NVIDIA-Open-Model-License (C-RADIO weights; commercial use allowed)
- loader-code
- caveat (the repository and the model code on the Hub are under the NVIDIA Source Code License for RADIO, research or evaluation use only)
The C-RADIO weights are under the NVIDIA Open Model License, which allows commercial use. The code that loads them, in the repository and on the Hub, is under NVIDIA's Source Code License for RADIO, which allows research or evaluation only. NVIDIA has released neither the 700-million-image training set nor the distillation training code.
- https://huggingface.co/api/models?author=nvidia&search=RADIO&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
nvidia C-RADIO, C-RADIOv2, v3 and v4 repositories gated false with license_name nvidia-open-model-license; RADIO-B, L, H and g tagged nvidia-source-code-license.
- https://huggingface.co/nvidia/C-RADIOv4-H/raw/main/radio_model.py recorded 2026-11-12
radio_model.py in the C-RADIOv4-H repository opens with "Copyright (c) 2023-2024, NVIDIA CORPORATION. All rights reserved." and reserves all rights absent "an express license agreement from NVIDIA".
- https://huggingface.co/nvidia/C-RADIOv4-H/raw/main/README.md recorded 2026-11-12
C-RADIOv4-H card: license_name nvidia-open-model-license; "This model is ready for commercial/non-commercial use"; Training Dataset "NV-CC-Img-Text-Dataset", "700 Million Images".
- https://raw.githubusercontent.com/NVlabs/RADIO/main/LICENSE recorded 2026-11-12
LICENSE: "NVIDIA Source Code License for RADIO"; 3.3 Use Limitation: "The Work and any derivative works thereof only may be used or intended for use non-commercially", meaning "for research or evaluation purposes only".
- https://raw.githubusercontent.com/NVlabs/RADIO/main/README.md recorded 2026-11-12
README Licensing: "Models prefixed with C-RADIO are governed by the NVIDIA Open Model License, which enables commercial use cases. Models prefixed with E-RADIO and RADIO are governed by the NSCL LICENSE file, which is non-commercial."; C-RADIOv4-SO400M and C-RADIOv4-H "Commercially Permissive".
- https://ungh.cc/repos/NVlabs/RADIO/files/main recorded 2026-11-12
Repository tree: radio/ model and adaptor code, examples for kNN and zero-shot ImageNet evaluation, and an mmseg probe; no distillation training script.
Adoption
2 medium confidenceHugging Face downloads over the trailing 30 days for the two C-RADIOv4 checkpoints and the three most-used C-RADIOv3 ones. Loads through PyTorch Hub fetch the same files and are counted; the non-commercial RADIO and E-RADIO models are not.
- https://huggingface.co/api/models?author=nvidia&search=RADIO&limit=100&expand[]=downloads&expand[]=cardData&expand[]=gated recorded 2026-11-12
C-RADIOv4-SO400M 15,208, C-RADIOv4-H 7,991, C-RADIOv3-L 4,012, C-RADIOv3-B 3,978 and C-RADIOv3-H 3,318 downloads in the trailing 30 days (34,507 together).
Capability
4 medium confidenceLike DINOv3, C-RADIO is mainly a backbone that other models build on, and its SigLIP2 adaptor classifies images against whatever labels the user types with no training. Dense tasks such as segmentation still need a trained head.
- https://huggingface.co/nvidia/C-RADIOv4-H/raw/main/README.md recorded 2026-11-12
C-RADIOv4-H card benchmark table: ImageNet zero-shot 83.09, kNN 86.68, ADE20k semantic segmentation 55.20 mIoU; teachers SigLIP2-g, DINOv3-7B and SAM3.
- https://raw.githubusercontent.com/NVlabs/RADIO/main/README.md recorded 2026-11-12
README: C-RADIOv4 adaptors "siglip2-g, dino_v3, sam3"; "The clip and siglip adaptors have the additional functionality of supporting tokenization and language encoding. Refer to examples/zero_shot_imagenet.py".
Verified 2026-11-12