Meta's vision model that sees forests and street markets without fine-tuning

What it does

DINOv3 provides pretrained vision transformers and ConvNeXt backbones that output dense, high-resolution features for images. The models come in sizes from 21M to 6.7B parameters, trained on either web-scale data (LVD-1689M) or satellite imagery (SAT-493M). Meta ships reference PyTorch code plus adapters for linear probing on tasks like semantic segmentation (ADE20K), depth estimation (NYUv2-Depth), and canopy height mapping.

The interesting bit

The pitch is “without fine-tuning” — these are foundation models in the original sense, meant to work as frozen feature extractors. The CHMv2 release is a nice flex: a 7B-parameter ViT pretrained on satellite data, repurposed for global forest canopy height mapping, with weights on Hugging Face and integration into the Transformers library.

Key highlights

• ViT and ConvNeXt variants from tiny (21M) to 7B parameters, all trained with self-supervised distillation
• Two pretraining domains: general web images (LVD-1689M) and satellite imagery (SAT-493M)
• Supported by PyTorch Hub, Hugging Face Transformers (≥4.56.0), and timm (≥1.0.20)
• Released task code: linear segmentation, depth estimation, and canopy height inference
• Model weights require an access request via Meta’s download portal; wget recommended over browser downloads

Caveats

• Weight downloads are gated behind a request form, not directly fetchable
• The README is heavy on model release announcements and light on training methodology or architecture details

Verdict

Worth a look if you need strong frozen visual features and don’t want to fine-tune a CLIP variant. Skip if you were hoping for an open-weights, no-gatekeeping drop-in replacement — the access friction is real.