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
- Paper RADIOv2.5: Improved Baselines for Agglomerative Vision Foundation Models CVPR 2025 · first released 10 Dec 2024
- Paper AM-RADIO: Agglomerative Vision Foundation Model -- Reduce All Domains Into One CVPR 2024 · first released 10 Dec 2023
- Paper PHI-S: Distribution Balancing for Label-Free Multi-Teacher Distillation arXiv · first released 2 Oct 2024
Paper 2024 Added in 0.2.0 Evaluated with 0.2.0
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.
| Backend | dtype | Time per call (ms) | vs original fp16 | Peak memory allocated / reserved | Load (s) | First call (s) |
|---|---|---|---|---|---|---|
| Original fp32hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0 | fp32 | 6.31 | 0.91× | 415 MiB/466 MiB | – | 0.17 |
| Original fp16torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0 | fp16 | 5.75 | – | 589 MiB/626 MiB | – | 0.14 |
| Portable fp32 | fp32 | not measured | – | – | – | – |
| Portable | fp16 | 3.04 | 1.9× | 217 MiB/246 MiB | – | 0.08 |
| Kernels | fp16 | 1.36 | 4.2× | 214 MiB/248 MiB | 1.7 | 0.06 |
| Portable + torch.compile | 7.33 | 0.78× | 278 MiB/306 MiB | 8 (build) | – | |
| TensorRT fp16 | fp16 | 1.19 | 4.8× | 230 MiBenginenot comparable | 37 (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
| Device | Backend | dtype | ms / call | Peak memory | Output vs reference | Input |
|---|---|---|---|---|---|---|
| Apple M1 Pro 16 GB MPS, torch 2.14.1, 2026-11-13 | Original | fp32 | 21.7 | 1,841 MiB process | – | 224x224 |
| Portable | fp16 | 16.5 | 2,230 MiB process | matches | 224x224 | |
| Portable | fp32 | 19.2 | 2,069 MiB process | matches | 224x224 | |
| Kernels | fp16 | not 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 Facefastforks 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)
| Run | dtype | top-1 % ↑ | Δ top-1 % vs fp32 original | images | ms / image |
|---|---|---|---|---|---|
| Original fp32 | fp32 | 57.58 | reference | 5000 | 3.80 |
| Original fp16 | fp16 | 57.62 | +0.04 | 5000 | 1.24 |
| Portable fp32 | not measured | ||||
| Portable | fp16 | 57.60 | +0.02 | 5000 | 1.02 |
| Kernels | fp16 | 57.66 | +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).
| Check | Tolerance | Measured |
|---|---|---|
| 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=falsefp32_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 classifiersfvm_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.
| variant | input | backend | time (ms) | peak memory (MiB, allocated / reserved) |
|---|---|---|---|---|
c-radio-v3-b | synthetic (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-l | synthetic (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-h | synthetic (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-g | synthetic (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-so400m | synthetic (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-h | synthetic (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-b3000x4000 (1.7 GiB peak allocated, 2.0 reserved);c-radio-v3-l3000x4000 (2.5 GiB peak allocated, 2.8 reserved);c-radio-v3-h3000x4000 (3.5 GiB peak allocated, 3.9 reserved);c-radio-v3-g3000x4000 (4.8 GiB peak allocated, 5.4 reserved);c-radio-v4-so400m3000x4000 (2.7 GiB peak allocated, 3.3 reserved);c-radio-v4-h3000x4000 (3.4 GiB peak allocated, 4.0 reserved) c-radio-v3-b224x224, first call: portable + torch.compile 8 s (cold Inductor cache), TensorRT fp16 37 s (ONNX + engine).c-radio-v3-l224x224, first call: portable + torch.compile 11 s (cold Inductor cache), TensorRT fp16 42 s (ONNX + engine).c-radio-v3-h224x224, first call: portable + torch.compile 15 s (cold Inductor cache), TensorRT fp16 45 s (ONNX + engine).c-radio-v3-g224x224, first call: portable + torch.compile 17 s (cold Inductor cache), TensorRT fp16 57 s (ONNX + engine).c-radio-v4-so400m224x224, first call: portable + torch.compile 12 s (cold Inductor cache), TensorRT fp16 49 s (ONNX + engine).c-radio-v4-h224x224, first call: portable + torch.compile 13 s (cold Inductor cache), TensorRT fp16 51 s (ONNX + engine).c-radio-v3-boriginal (fp32) 224x224: hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0c-radio-v3-boriginal 224x224: torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0c-radio-v3-loriginal (fp32) 224x224: hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0c-radio-v3-loriginal 224x224: torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0c-radio-v3-horiginal (fp32) 224x224: hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0c-radio-v3-horiginal 224x224: torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0c-radio-v3-goriginal (fp32) 224x224: hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0c-radio-v3-goriginal 224x224: torch.autocast(fp16) around the forward; hf_model.RADIOModel (trust_remote_code) summary, transformers 5.19.0c-radio-v4-so400moriginal 224x224: NVIDIA RADIOModel (trust_remote_code) summaryc-radio-v4-horiginal 224x224: NVIDIA RADIOModel (trust_remote_code) summaryc-radio-v3-bload / first call: original – s / 0.145 s, portable – s / 0.083 s, kernels – s / 0.06 sc-radio-v3-lload / first call: original – s / 0.154 s, portable – s / 0.071 s, kernels – s / 0.038 sc-radio-v3-hload / first call: original – s / 0.164 s, portable – s / 0.07 s, kernels – s / 0.071 sc-radio-v3-gload / first call: original – s / 0.171 s, portable – s / 0.079 s, kernels – s / 0.075 sc-radio-v4-so400mload / first call: original 3.42 s / 0.201 s, portable 2.43 s / 0.067 s, kernels 2.52 s / 0.072 sc-radio-v4-hload / 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.
| 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
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).
| variant | kernels error vs original, mean / p99 | GPU, torch | kernels used |
|---|---|---|---|
c-radio-v3-b | 0.00566 / 0.02299 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_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-l | 0.03281 / 0.14863 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_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-h | 0.00382 / 0.01776 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_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-g | 0.00683 / 0.03369 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_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-so400m | 0.00571 / 0.02814 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_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-h | 0.00377 / 0.0154 | RTX 5060 Ti, torch 2.14.1+cu130 | cudnn_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 = trueputs 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: Chelsea the cat, Stéfan van der Walt (CC0).