C-RADIOv3 / v4 (fastforks)

C-RADIOv3 and C-RADIOv4 (NVIDIA 2025 / 2026), the commercially licensed agglomerative ViTs: v3 distilled from CLIP, SigLIP 2, DINOv2 and SAM (ViT-B / L / H / g), v4 from SigLIP 2-g, DINOv3-7B and SAM 3 (ViT-SO400M / H); patch 16, dense patch tokens and a summary embedding (v3: 3 teachers, v4: 2) at any size up to 2048 px.

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://huggingface.co/nvidia/C-RADIOv3-B at 44653a0482cf.

Author weight repos: https://huggingface.co/nvidia/C-RADIOv3-B, https://huggingface.co/nvidia/C-RADIOv3-H, https://huggingface.co/nvidia/C-RADIOv3-L, https://huggingface.co/nvidia/C-RADIOv3-g, https://huggingface.co/nvidia/C-RADIOv4-H, https://huggingface.co/nvidia/C-RADIOv4-SO400M.

Task: features. Tier: experimental. Outputs: last_hidden_state, pooler_output, hidden_states. Backends: portable, kernels.

Input: one RGB image. Output: dense patch features and a pooled image feature.

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant weights (author link) SHA-256 weights licence commercial use
c-radio-v3-b@1 c-radio-v3-b.safetensors 27881c65f10a LicenseRef-NVIDIA-Open-Model-License yes
c-radio-v3-g@1 c-radio-v3-g.safetensors 8234e4556fcb LicenseRef-NVIDIA-Open-Model-License yes
c-radio-v3-h@1 c-radio-v3-h.safetensors 58d6b8cc6907 LicenseRef-NVIDIA-Open-Model-License yes
c-radio-v3-l@1 c-radio-v3-l.safetensors 2aedceebd3f8 LicenseRef-NVIDIA-Open-Model-License yes
c-radio-v4-h@1 c-radio-v4-h.safetensors 851f28346811 LicenseRef-NVIDIA-Open-Model-License yes
c-radio-v4-so400m@1 c-radio-v4-so400m.safetensors 23e0c117de49 LicenseRef-NVIDIA-Open-Model-License yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0.

Training data: NV-CC-Img-Text-Dataset (NVIDIA web image-text data, 700 M images, not released) (700 M web images (NV-CC-Img-Text-Dataset), distilled from teachers whose own training data does not carry over as terms: v3 the DFN CLIP ViT-H/14, SigLIP 2 (So400m; g for C-RADIOv3-H), DINOv2-g and SAM ViT-H; v4 SigLIP 2-g, DINOv3-7B and SAM 3).

  • Licence note: The weights are governed by the NVIDIA Open Model License Agreement, which allows commercial use (the model card: 'ready for commercial/non-commercial use'). Our code is timm's VisionTransformer plus our own walk; no NVIDIA code is vendored.

Licence evidence: https://huggingface.co/nvidia/C-RADIOv3-B/blob/44653a0482cf460bb4f12595fc3cc3dfecc403d1/README.md and https://huggingface.co/nvidia/C-RADIOv4-SO400M/blob/c0457f5dc26ca145f954cd4fc5bb6114e5705ad8/README.md

Input size

any H x W >= 1, resized or padded inside the model Tested up to (portable, fp16, RTX 5060 Ti 16 GB): c-radio-v3-b 3000x4000 (1.7 GiB peak allocated, 2.0 reserved); c-radio-v3-g 3000x4000 (4.8 GiB peak allocated, 5.4 reserved); c-radio-v3-h 3000x4000 (3.5 GiB peak allocated, 3.9 reserved); c-radio-v3-l 3000x4000 (2.5 GiB peak allocated, 2.8 reserved); c-radio-v4-h 3000x4000 (3.4 GiB peak allocated, 4.0 reserved); c-radio-v4-so400m 3000x4000 (2.7 GiB peak allocated, 3.3 reserved)

Speed and memory

Features, synthetic, 224x224

model time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
c-radio-v3-b 6.31 5.75 3.04 1.36 4.2x 2.2x 589 / 626 217 / 246 214 / 248
c-radio-v3-g 58.0 29.5 14.6 10.7 2.8x 1.4x 6672 / 6804 2266 / 2308 2263 / 2310
c-radio-v3-h 34.5 17.0 8.90 7.43 2.3x 1.2x 3810 / 4110 1289 / 1308 1286 / 1308
c-radio-v3-l 16.8 10.7 5.53 3.28 3.3x 1.7x 1838 / 1878 648 / 686 645 / 706
c-radio-v4-h – 17.1 8.92 7.41 2.3x 1.2x 6214 / 6250 1289 / 1308 1289 / 1308
c-radio-v4-so400m – 11.5 6.33 4.25 2.7x 1.5x 4115 / 4150 864 / 906 864 / 908

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
c-radio-v3-b – – – 0.14 0.08 0.06
c-radio-v3-g – – – 0.17 0.08 0.07
c-radio-v3-h – – – 0.16 0.07 0.07
c-radio-v3-l – – – 0.15 0.07 0.04
c-radio-v4-h 5.55 6.43 2.77 0.20 0.07 0.07
c-radio-v4-so400m 3.42 2.43 2.52 0.20 0.07 0.07

torch.compile and TensorRT

model input time: portable compile TensorRT fp16 kernels preparation: compile (cold first call) ONNX + engine build peak MiB (alloc / reserved): portable compile kernels TensorRT (engine + torch)
c-radio-v3-b 224x224 3.04 ms 7.33 ms 1.19 ms 1.36 ms 8.29 s 37.4 s 217 / 246 278 / 306 214 / 248 427 / 454
c-radio-v3-g 224x224 14.6 ms 25.5 ms 10.3 ms 10.7 ms 16.5 s 57.1 s 2266 / 2308 2329 / 2368 2263 / 2310 4483 / 4530
c-radio-v3-h 224x224 8.90 ms 18.1 ms 6.62 ms 7.43 ms 14.7 s 44.7 s 1289 / 1308 1350 / 1386 1286 / 1308 2558 / 2586
c-radio-v3-l 224x224 5.53 ms 11.8 ms 3.06 ms 3.28 ms 11.3 s 41.8 s 648 / 686 710 / 746 645 / 706 1297 / 1350
c-radio-v4-h 224x224 8.92 ms 17.9 ms 6.64 ms 7.41 ms 12.9 s 51.2 s 1289 / 1308 1351 / 1388 1289 / 1308 2558 / 2586
c-radio-v4-so400m 224x224 6.33 ms 14.2 ms 3.75 ms 4.25 ms 11.5 s 49.0 s 864 / 906 927 / 966 864 / 908 1695 / 1740

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

On public datasets

Two disjoint seeded subsets of the val folders; each image shorter side 256 (bicubic), centre crop 224, the CLS / pooled feature; cosine kNN, vote weighted by similarity. Original = the original code named under Details, same checkpoint and protocol.

variant dataset original fp32 original fp16 portable kernels note
c-radio-v3-b ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) 57.58 57.62 57.60 57.66
c-radio-v3-l ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) 67.06 67.06 67.08 67.08
c-radio-v3-h ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) 69.52 69.52 69.54 69.58
c-radio-v3-g ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) 70.88 70.94 70.92 70.98
c-radio-v4-so400m ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) 70.06 70.02 70.02 70.04
c-radio-v4-h ImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images) 71.38 71.36 71.32 71.36

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks 28db82cb / 8bce3316 / cf1956ec, 2026-11-11 to 2026-11-13.

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.0001 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 5 % and p99 <= 23 % of the reference's median |value| (absolute error of the output last_hidden_state).

Limitations

  • last_hidden_state holds the prefix tokens (v3: 4 summary + 4 registers; v4: 3 summary + 7 registers), then the patches (upstream's features).
  • pooler_output is upstream's summary: the summary tokens concatenated (v3: clip, siglip2, dino_v2 = 3 x hidden size; v4: siglip2, dino_v3 = 2 x hidden size).
  • The teacher adaptors (the heads mapping the summary to CLIP / SigLIP 2 text space) are not included.
  • v4 is tested by NVIDIA from 256 to 2048 px; our goldens check it at 112 px like v3 (the same code path).
  • kernels replays a CUDA graph per small input shape without hidden states (up to about 1M activations: ViT-B ~460 px, ViT-L ~300 px), captured on the shape's second call, at most 4 per model; model.release_graphs() frees them.
  • Small kernels calls (graph-sized, with or without layers) run GEMMs of at most 600 tokens on cuBLAS: faster in a graph replay (ViT-B/16 224 px: 1.26 against 1.52 ms), about 10 % slower eagerly on ViT-T / B; a replay equals the eager call bit for bit.
  • fp32_accumulate = true puts every GEMM on cuBLAS with fp32 accumulation (the portable precision; slower on GeForce).
  • Kernels at 224 x 224 peak above portable for c-radio-v3-b (281 vs 214 MB; original 589 MB), about the size of one extra fc2 copy (57 MB) plus ~30 MB of CUDA-graph capture; the extra copies are not shared with the original parameters.

Usage

from fastforks import features

model = features.load("c-radio-v3-b")  # mode="auto": the preferred backend from recorded measurements
result = model("photo.jpg", size=518)
result.last_hidden_state  # tokens [N, C], as Hugging Face's output
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: c-radio-v3-b, c-radio-v3-g, c-radio-v3-h, c-radio-v3-l, c-radio-v4-h, c-radio-v4-so400m.

Kernels backend

features.load("c-radio-v3-b", mode="kernels") runs our CUDA kernels (cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.

Downloads last month

-

Downloads are not tracked for this model. How to track
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for fastforks/c-radio

Finetuned
(3)
this model

Collection including fastforks/c-radio