C-RADIOv4 Vision Backbone Models

• C-RADIOv4 is a fourth-generation vision backbone employing a refined ViT architecture with any-resolution support and efficient computation.
• It integrates multi-teacher distillation from SigLIP2, DINOv3, and SAM3 to enhance dense prediction and open-set segmentation capabilities.
• Available in SO400M and H variants, it balances global feature representation with computational efficiency to achieve competitive performance on dense tasks.

C-RADIOv4 is the fourth-generation release from the C-RADIO family of agglomerative vision backbone models, designed to unify and advance the distinct capabilities of multiple teacher networks through multi-teacher distillation. Building on AM-RADIO/RADIOv2.5 architectures, C-RADIOv4 applies a refined ViT-style backbone and incorporates innovations for efficient computation, any-resolution input, and enhanced downstream task performance, while maintaining computational parity with previous releases. The family consists of two primary variants—SO400M (412 million parameters) and H (631 million parameters)—each trained on a curated set of leading teacher models: SigLIP2, DINOv3, and SAM3. C-RADIOv4 models offer improved representational power, support open-set segmentation via SAM3 imitation, and are distributed under a permissive Apache 2.0 license (Ranzinger et al., 24 Jan 2026).

1. Model Variants and Parameterization

C-RADIOv4 encompasses two configurations: SO400M ("small/optimal"; 412M parameters) and H ("huge"; 631M parameters). Both retain the plain-ViT backbone topology derived from RADIOv2.5 but are differentiated by the following hyperparameters:

Variant	Embedding Dim (approx)	Transformer Layers	MLP Width	Parameters
SO400M	≈768	12–16	Proportional to embed	412M
H	1,024–1,280	24–32	≈4× embedding dim	631M

SO400M, with 35% fewer parameters than H, matches or exceeds previous ViT-H quality on many dense tasks, yielding efficiency without significant performance compromise. H remains the performance leader within the family.

2. Architectural Features and Efficiency Enhancements

The C-RADIOv4 backbone adheres to plain-ViT principles with several design refinements:

• Patch Embedding: Linear projection of 16×16 pixel patches into token space, mirroring standard ViT [Dosovitskiy et al. 2021].
• Positional Encoding: Fixed 2D sine-cosine embeddings shared across all input resolutions for true any-resolution support.
• Transformer Blocks: Combination of Multi-Head Self-Attention (MHSA), MLP, LayerNorm, and residual connections. Four of these layers are sparsely global MHSA blocks, with the remainder operating in windowed mode; window size is adjustable for ViTDet operation.
• Activation and Output: GELU activations in the MLP. Attention-output projections employ full embedding dimensionality (no bottleneck).
• Efficiency Trade-off: Windowed attention (e.g., 8×8 or 12×12 windows) reduces the quadratic cost $O(T^2)$ to $O(T \cdot w^2)$, retaining global context via the occasional global MHSA layers.

This architectural composition allows C-RADIOv4 to balance high-resolution efficiency against global representational capacity, particularly when ViTDet mode is enabled.

3. Multi-Teacher Distillation Mechanism

The distillation strategy employs feature and logit matching across three teacher models (SigLIP2, DINOv3, SAM3), using a weighted loss function comprising spatial and summary-token components. The summary (soft-teacher) loss is defined by:

$$L_\text{distill} = \sum_{i \in \{\text{SigLIP2}, \text{DINOv3}, \text{SAM3}\}} \alpha_i \cdot D_{KL}\bigl(\sigma(T_i(x)/T) \;||\; \sigma(S(x)/T)\bigr)$$

Where $T_i(x)$ are teacher logits, $S(x)$ is the student summary token, $\sigma(\cdot)$ is softmax, temperature $T = 0.07$, and the teacher weights $\alpha_i = 1/3$.

Additional loss terms include:

• Spatial Feature Loss ($L_\text{spatial}$): Match 2D dense maps with PHI-S normalization and randomized patch shifts to enforce shift-equivariance.
• Balanced Angle Loss ($L_\text{angle}$): Normalize cone radii of summary-tokens across teachers to prevent collapse.
• Regularization: MESA [Du et al. 2022] and DAMP [Trinh et al. 2024] enhance flat minima and robustness to noise.

Incorporation of SAM3 as a teacher grants the ability to synthesize features actionable by SAM3’s text-and-point driven mask decoder, subsequently enabling open-set segmentation analogous to the Segment Anything Model paradigm.

4. Any-Resolution Support and Training Protocol

C-RADIOv4 introduces stochastic resolution sampling per batch, rather than fixed input sizes:

• Low-resolution grid: 128, 192, 224, 256, 384, 432 px
• High-resolution grid: 512, 768, 1024, 1152 px

SigLIP2 embeddings utilize FeatSharp upsampling to 3× (384→1152 px), while SAM3 operates on 1152×1152 px with mosaic augmentation. This diverse sampling yields monotonically smooth performance scaling across resolutions. Notably:

Resolution (px)	ADE20k IoU (C-RADIOv4-H)	DINOv3-7B
512	55.20	55.9
1024	57.02	57.3
1536	57.72	57.8

This suggests C-RADIOv4-H matches the quality of significantly larger 7B-parameter SSL backbones at a fraction of the parameter count.

5. ViTDet-Mode: High-Resolution Efficiency

Building on ViTDet [Li et al. 2022], a core operational mode for C-RADIOv4 assigns nearly all transformer layers to local-window attention, retaining only four global MHSA blocks. Efficiency and performance metrics under this regime include:

Model/Window	Latency (1152×1152, ms)	COCO box AP	COCO mask AP
SAM3 (ViT-L+, global)	58	—	—
SO400M (W=8)	43	—	—
H (W=12)	45	61.3	52.7
ViT-H+FPN (global)	—	60.8	52.0

ViTDet-enabled inference achieves up to ~25% faster execution than SAM3 at 1152² resolution, with only minor AP penalty as window size decreases.

6. Downstream Task Performance

C-RADIOv4 models demonstrate competitive or superior results across multiple vision tasks. Relevant benchmarks include:

Task	RADIOv2.5-H	C-RADIOv3-H	SO400M	H
ImageNet-1K Zero-Shot	82.51	82.65	82.01	83.09
k-NN (256 px)	—	—	—	86.59
Semantic Segmentation (ADE20k IoU)	51.58	52.75	55.14	55.20
Probe3D (Depth/Surface/NAVI/SPair)	85.69/62.46/60.89/56.24	—	85.29/61.91/62.44/60.01	85.55/61.70/63.44/60.57
SA-Co/Gold Instance Segmentation (cgF1 avg)	—	—	—	44.7 (global), 43.7 (W=8)

A plausible implication is that multi-teacher distillation and architectural refinements substantially boost dense prediction and 3D probing performance versus prior backbone models.

7. Computational Complexity and Licensing

Computational efficiency profiles for C-RADIOv4 variants are summarized below:

Resolution (px)	SO400M FLOPs	H FLOPs	SO400M Latency (ms)	H Latency (ms)
224×224 (≈14×14)	40 GFLOPs	80 GFLOPs	12	22
1024×1024 (64×64)	670 GFLOPs	1.3 TFLOPs	90	180
4096×4096	—	—	400	950

With ViTDet windowed attention, computational complexity transitions from quadratic to approximately linear scaling with token count for large $O(T \cdot w^2)$0.

C-RADIOv4 is distributed under the permissive Apache 2.0 license, enabling unrestricted commercial and academic use, fine-tuning, and redistribution, provided attribution and NOTICE file retention.

---

C-RADIOv4 delivers ViT-style backbones with any-resolution support, high efficiency for high-resolution inference, and multi-teacher distillation leveraging SigLIP2, DINOv3, and SAM3. The family achieves state-of-the-art results on classification, dense pixel prediction, and open-set segmentation tasks at reduced compute and parameter scale relative to large SSL backbones, and is available for open use (Ranzinger et al., 24 Jan 2026).