NEWVectors or files. Pick a path.Start →
    Models/Embeddings/facebook/dinov3-large
    PyTorchVisual EmbeddingsApache 2.0

    dinov3-large

    by facebook

    Next-generation self-supervised vision model with Gram anchoring and 6.7B scaling

    450Kdl/month
    300M (Large), 6.7B (ViT-7B)params
    Identifiers
    Model ID
    facebook/dinov3-large
    Feature URI
    mixpeek://image_extractor@v1/facebook_dinov3_large_v1

    Overview

    DINOv3 is Meta AI's successor to DINOv2, introducing Gram anchoring to solve dense feature degradation during long training schedules. It scales up to 6.7B parameters (ViT-7B) and trains on 1.7 billion web images plus 493M satellite images, making it the most versatile vision foundation model available.

    On Mixpeek, DINOv3 delivers state-of-the-art visual features for tasks ranging from classification and segmentation to satellite/aerial imagery analysis, all without fine-tuning.

    Architecture

    Vision Transformer with patch size 16. Scales from ViT-S (21M) to ViT-7B (6.7B params). Introduces Gram anchoring to stabilize dense features during extended training. Also distills into ConvNeXt backbones. Supports flexible resolution and post-hoc text alignment.

    Mixpeek SDK Integration

    // No extractor parameter takes a Hugging Face model id (checked against
    // GET /v1/discovery/extractors, which returns 13), so dinov3-large runs
    // on your side and the output is upserted through POST
    // /v1/namespaces/{namespace_id}/documents/upsert. On Enterprise the other
    // path is to upload the weights instead: POST /v1/namespaces/{id}/models
    // accepts the huggingface format and a custom plugin loads them.
    const res = await fetch(
      "https://api.mixpeek.com/v1/namespaces/ns_your_namespace/documents/upsert",
      {
        method: "POST",
        headers: {
          Authorization: "Bearer API_KEY",
          "Content-Type": "application/json",
        },
        body: JSON.stringify({
          collection_id: "col_your_collection",
          documents: [
            {
              document_id: "asset-00412",
              // The vector name has to match a vector index on the collection.
              vectors: { "image-embedding": yourVector },
              payload: { source_key: "archive/2026/asset-00412" },
            },
          ],
        }),
      },
    );
    
    // Managed alternative, if this exact model is not the requirement:
    // image_extractor@v1 runs google/siglip-base-patch16-224
    // (768-d) over a bucket, with no inference of your own.

    Capabilities

    • Gram anchoring for stable dense feature training
    • Scales up to 6.7B parameters (ViT-7B)
    • Trained on 1.7B web + 493M satellite images
    • ViT and ConvNeXt backbone variants
    • Multi-domain: natural images and satellite/aerial imagery

    Use Cases on Mixpeek

    High-fidelity visual search across massive image collections
    Satellite and aerial imagery analysis
    Dense segmentation and depth estimation
    Foundation for downstream classification without fine-tuning

    Benchmarks

    DatasetMetricScoreSource
    ImageNet (linear probe)Top-1 Accuracy83.1%DINOv3 model card

    Performance

    Input Size224×224 px
    Embedding Dim1024
    GPU Latency~11ms / image (A100)
    CPU Latency~130ms / image
    GPU Throughput~90 images/sec (A100)
    GPU Memory~1.3 GB

    Specification

    FrameworkPyTorch
    Organizationfacebook
    FeatureVisual Embeddings
    Output768-dim vector
    Modalitiesvideo, image
    RetrieverVector Search
    Parameters300M (Large), 6.7B (ViT-7B)
    LicenseApache 2.0
    Downloads/mo450K

    Research Paper

    DINOv3

    arxiv.org

    Build a pipeline with dinov3-large

    Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

    Run it on your own data, free