Models · Feature backbones

C-RADIOv3 / v4

NVIDIA's C-RADIOv3 (B, L, H, g): one backbone distilled from several teachers (DINOv2, SigLIP 2, SAM), image features at any resolution. NVIDIA Open Model License.

experimental Fork of https://huggingface.co/nvidia/C-RADIOv3-B@44653a0 · backends: portable, kernels

Checkpoint

Paper 2024 Added in 0.2.0 Evaluated with 0.2.0

Kernels, RTX 5060 Ti 16 GB 1.36 ms 224x224, fp16
vs the faster original 4.2× original fp16: 5.75 ms
Peak memory 214 MiB original: 589 MiB
Accuracy vs original fp32 +0.08 kernels top-1 %, ImageNet-1k val (57.58 original)

Use it

pip install fastforks

Lean install with only what this model needs: pip install "fastforks-core[timm]"

from fastforks import features

model = features.load("c-radio-v3-b")  # kernels on a supported GPU, else portable PyTorch
result = model(image)
result.last_hidden_state  # also: pooler_output, hidden_states
result.backend  # what ran: mode, device, dtype, kernels
fastforks bench c-radio-v3-b

Times the model on this machine. load(…, mode="portable") or mode="kernels" picks a backend; usage guide.

Speed and memory

Every backend on one GPU (RTX 5060 Ti 16 GB) and the same input, the authors’ code included. Kernels runs fp16. Portable runs fp16 on CUDA.

BackenddtypeTime per call (ms)vs original fp16Peak memory
allocated / reserved
Load (s)First call (s)
Original fp32hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0fp32 6.310.91×415 MiB/466 MiB–0.17
Original fp16torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0fp16 5.75–589 MiB/626 MiB–0.14
Portable fp32fp32not measured ––––
Portablefp16 3.041.9×217 MiB/246 MiB–0.08
Kernelsfp16 1.364.2×214 MiB/248 MiB1.70.06
Portable + torch.compile 7.330.78×278 MiB/306 MiB8 (build)–
TensorRT fp16fp16 1.194.8×230 MiBenginenot comparable37 (build)–

Measured on RTX 5060 Ti 16 GB, torch 2.14.1+cu130, 2026-11-10 to 2026-11-15. First call, 3 warm-up, peak reset, median of 3 rounds x median of 5. Ratios divide two times of this table (same GPU, same input). Original fp16 is the authors’ own precision option, else torch.autocast. torch.compile and TensorRT: build time instead of load time. TensorRT memory (engine, not comparable) is the GPU memory the engine took, read from the driver because TensorRT allocates outside PyTorch, so it can't be compared with the PyTorch peaks of the other rows.

Other devices

DeviceBackenddtypems / callPeak memoryOutput vs referenceInput
Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-13Originalfp3221.7 1,841 MiB process–224x224
Portablefp1616.5 2,230 MiB processmatches224x224
Portablefp3219.2 2,069 MiB processmatches224x224
Kernelsfp16not implemented–224x224

CUDA memory is torch’s peak allocated / reserved; Apple has no such counter, so the Mac rows show the process footprint (weights, Python and torch included). “Output vs reference” compares the device’s output with the stored golden output under the card’s tolerance; – = not checked on that device. Kernels “not implemented” = no kernels for this model on that backend yet: it runs portable there.

Licence and weights

Commercial use allowed

Code
Apache-2.0
Weights
LicenseRef-NVIDIA-Open-Model-License source
Training data
  • 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.

Our reading of the licences, not legal advice.

Weights

c-radio-v3-b.safetensors 393 MB· hosted on Hugging Face

fastforks downloads and checks these files on first use (fastforks download c-radio-v3-b).

Accuracy

The test set through the authors’ code and through ours: same checkpoint, same protocol. The distance between the original’s fp32 and fp16 runs is what half precision alone costs; our runs should stay within it.

ImageNet-1k val

kNN top-1 % (k = 20, 10000 labelled / 5000 test images)

Rundtypetop-1 % ↑Δ top-1 % vs fp32 originalimagesms / image
Original fp32fp3257.58 reference 5000 3.80
Original fp16fp1657.62 +0.04 5000 1.24
Portable fp32not measured
Portablefp1657.60 +0.02 5000 1.02
Kernelsfp1657.66 +0.08 5000 0.638

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 authors’ code, same checkpoint and protocol. Measured on original fp32, original fp16: RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-12; portable, kernels: RTX 5060 Ti, torch 2.14.1+cu130, fastforks 8bce3316, 2026-11-11.

Closeness to the original

Beyond the test-set scores: the raw output of ours against the original’s fp32 output on the same input, pixel by pixel. 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).

CheckToleranceMeasured
CPU fp32, max |diff|≤ 0.0001 (relative)not recorded
cpu_rtol≤ 0.00002 (relative)not recorded
GPU default dtypes, mean |diff|≤ 0.05 (relative)0.00566 (relative)
GPU default dtypes, p99 |diff|≤ 0.23 (relative)0.02299 (relative)

Recorded for c-radio-v3-b on RTX 5060 Ti, torch 2.14.1+cu130, 2026-11-10, with these kernels: cudnn_haccfvm_convfvm_extrafvm_mbfvm_r3

Details

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.

Input / output
Input: one RGB image. Output: dense patch features and a pooled image feature.
Input size
any H x W >= 1, resized or padded inside the model; tested up to 3000x4000 (1,712 MiB allocated)
Options
  • fp16_accumulate=false
  • fp32_accumulate=false
Outputs
last_hidden_state, pooler_output, hidden_states
Checkpoint
embed_dim=768, depth=12, num_heads=12, layer_scale=true, pos_align_corners=true
CUDA kernels
  • cudnn_haccconvolutions with fp16 accumulation (cuDNN)
  • fvm_convfused convolutions for image classifiers
  • fvm_extrafused transformer row operations (rotary split, SwiGLU with LayerNorm)
  • fvm_mbfused MBConv blocks (MobileNet and EfficientNet families)
  • fvm_r3further fused classifier blocks
Apple (MPS)
fp32, fp16

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.
Full reference, generated from the repository

NVIDIA's C-RADIOv3 (B, L, H, g): one backbone distilled from several teachers (DINOv2, SigLIP 2, SAM), image features at any resolution. NVIDIA Open Model License.

Commercial use: yes.

variantinputbackendtime (ms)peak memory (MiB, allocated / reserved)
c-radio-v3-bsynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
6.31
5.75
3.04
1.36
7.33
1.19
415 / 466
589 / 626
217 / 246
214 / 248
278 / 306
427 / 454
c-radio-v3-lsynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
16.8
10.7
5.53
3.28
11.8
3.06
1263 / 1278
1838 / 1878
648 / 686
645 / 706
710 / 746
1297 / 1350
c-radio-v3-hsynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
34.5
17.0
8.90
7.43
18.1
6.62
2595 / 2666
3810 / 4110
1289 / 1308
1286 / 1308
1350 / 1386
2558 / 2586
c-radio-v3-gsynthetic (224x224)original (fp32)
original
portable
kernels
portable + torch.compile
TensorRT fp16
58.0
29.5
14.6
10.7
25.5
10.3
4514 / 4554
6672 / 6804
2266 / 2308
2263 / 2310
2329 / 2368
4483 / 4530
c-radio-v4-so400msynthetic (224x224)original
portable
kernels
portable + torch.compile
TensorRT fp16
11.5
6.22
4.14
14.2
3.75
4115 / 4150
861 / 906
861 / 908
927 / 966
1695 / 1740
c-radio-v4-hsynthetic (224x224)original
portable
kernels
portable + torch.compile
TensorRT fp16
17.1
8.91
7.40
17.9
6.64
6214 / 6250
1286 / 1308
1286 / 1308
1351 / 1388
2558 / 2586
Details
  • Original code: https://huggingface.co/nvidia/C-RADIOv3-B at 44653a0482cf
  • Code licence: Apache-2.0 (permissive); files:
  • Weights licence: LicenseRef-NVIDIA-Open-Model-License
  • 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
  • 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.
  • Options: fp16_accumulate = False (precision only), fp32_accumulate = False (precision only)
  • Kernel libraries: cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3
  • Variants: c-radio-v3-b, c-radio-v3-l, c-radio-v3-h, c-radio-v3-g, c-radio-v4-so400m, c-radio-v4-h
  • Outputs: last_hidden_state, pooler_output, hidden_states
  • 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-l 3000x4000 (2.5 GiB peak allocated, 2.8 reserved); c-radio-v3-h 3000x4000 (3.5 GiB peak allocated, 3.9 reserved); c-radio-v3-g 3000x4000 (4.8 GiB peak allocated, 5.4 reserved); c-radio-v4-so400m 3000x4000 (2.7 GiB peak allocated, 3.3 reserved); c-radio-v4-h 3000x4000 (3.4 GiB peak allocated, 4.0 reserved)
  • c-radio-v3-b 224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 37 s (ONNX + engine).
  • c-radio-v3-l 224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 42 s (ONNX + engine).
  • c-radio-v3-h 224x224, first call: portable + torch.compile 15 s (cold Inductor cache), TensorRT fp16 45 s (ONNX + engine).
  • c-radio-v3-g 224x224, first call: portable + torch.compile 17 s (cold Inductor cache), TensorRT fp16 57 s (ONNX + engine).
  • c-radio-v4-so400m 224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 49 s (ONNX + engine).
  • c-radio-v4-h 224x224, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 51 s (ONNX + engine).
  • c-radio-v3-b original (fp32) 224x224: hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0
  • c-radio-v3-b original 224x224: torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0
  • c-radio-v3-l original (fp32) 224x224: hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0
  • c-radio-v3-l original 224x224: torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0
  • c-radio-v3-h original (fp32) 224x224: hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0
  • c-radio-v3-h original 224x224: torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0
  • c-radio-v3-g original (fp32) 224x224: hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0
  • c-radio-v3-g original 224x224: torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0
  • c-radio-v4-so400m original 224x224: NVIDIA RADIOModel (trust_remote_code) summary
  • c-radio-v4-h original 224x224: NVIDIA RADIOModel (trust_remote_code) summary
  • c-radio-v3-b load / first call: original – s / 0.145 s, portable – s / 0.083 s, kernels – s / 0.06 s
  • c-radio-v3-l load / first call: original – s / 0.154 s, portable – s / 0.071 s, kernels – s / 0.038 s
  • c-radio-v3-h load / first call: original – s / 0.164 s, portable – s / 0.07 s, kernels – s / 0.071 s
  • c-radio-v3-g load / first call: original – s / 0.171 s, portable – s / 0.079 s, kernels – s / 0.075 s
  • c-radio-v4-so400m load / first call: original 3.42 s / 0.201 s, portable 2.43 s / 0.067 s, kernels 2.52 s / 0.072 s
  • c-radio-v4-h load / first call: original 5.55 s / 0.195 s, portable 6.43 s / 0.071 s, kernels 2.77 s / 0.072 s
  • Protocol, first call, 3 warm-up, peak reset, median of 3 rounds x median of 5

Measured

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.

variantdatasetoriginal fp32original fp16portablekernelsnote
c-radio-v3-bImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)57.5857.6257.6057.66
c-radio-v3-lImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)67.0667.0667.0867.08
c-radio-v3-hImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)69.5269.5269.5469.58
c-radio-v3-gImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)70.8870.9470.9270.98
c-radio-v4-so400mImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)70.0670.0270.0270.04
c-radio-v4-hImageNet-1k val, kNN top-1 % (k = 20, 10000 labelled / 5000 test images)71.3871.3671.3271.36

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

Parity

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).

variantkernels error vs original, mean / p99GPU, torchkernels used
c-radio-v3-b0.00566 / 0.02299RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
c-radio-v3-l0.03281 / 0.14863RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
c-radio-v3-h0.00382 / 0.01776RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
c-radio-v3-g0.00683 / 0.03369RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
c-radio-v4-so400m0.00571 / 0.02814RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120
c-radio-v4-h0.00377 / 0.0154RTX 5060 Ti, torch 2.14.1+cu130cudnn_hacc@cudnn92400, fvm_conv@f1d064e539e1:sm_120, fvm_extra@3208ce525a91:sm_120, fvm_mb@720dd3f29bf7:sm_120, fvm_r3@f54eb890e23c:sm_120

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.

Example output

One fixed input through this checkpoint, the same for every model of the task (kernels · float16 · NVIDIA GeForce RTX 5060 Ti).

Input
Output
Chelsea the cat, input
Input · 100%

Input: Chelsea the cat, Stéfan van der Walt (CC0).