Release of ViT-S/16 distilled weights for the SAT-493M (Satellite) dataset

[isuchy]

Hi DINOv3 Team,

Thank you for the impressive release of DINOv3. The foundation models pretrained on satellite imagery (SAT-493M) are a significant contribution to the remote sensing community.

Currently, the SAT-493M models are available in two sizes:

ViT-L/16 distilled (300M parameters)
ViT-7B/16 (6,716M parameters)
While these are excellent for server-side processing, they are too resource-intensive for edge computing scenarios. I am specifically looking to deploy DINOv3 on NVIDIA Jetson devices (such as Orin Nano/NX/AGX) for real-time or on-site satellite data analysis.

On these edge platforms, memory (VRAM) and compute power are limited. The ViT-S/16 distilled variant (approx. 21M parameters), which is already available for the LVD-1689M (web) dataset, would be the perfect balance of performance and efficiency for Jetson-based applications.

Request:
Would it be possible to release a ViT-S/16 distilled version pretrained on the SAT-493M dataset?

Having a "Small" variant specialized for satellite imagery would enable:

Low-latency inference on edge hardware.
Reduced memory footprint, allowing for concurrent tasks on Jetson modules.
On-device processing for drones or remote monitoring stations where high-end server GPUs are not an option.
Thank you for your consideration and for supporting efficient AI deployment!

Best regards,
Ivan Suchy