vjepa2-vitg-fpc64-256

Name: vjepa2-vitg-fpc64-256
Author: facebook

by facebook

Highest-capacity V-JEPA 2 video encoder: self-supervised temporal representations

140Kdl/month

57likes

1.0Bparams

HuggingFace Run it on your own data, free

Identifiers

Model ID

facebook/vjepa2-vitg-fpc64-256

Feature URI

mixpeek://video_extractor@v1/facebook_vjepa2_vitg_fpc64_256_v1

Overview

V-JEPA 2 (ViT-g) is the largest checkpoint of Meta FAIR's video representation model. What makes the JEPA (Joint-Embedding Predictive Architecture) family different from a masked autoencoder is *where* it predicts: it masks spacetime regions of a clip and predicts the missing regions' **representations in latent space**, not their raw pixels. Skipping pixel reconstruction means the model never spends capacity on texture and lighting detail it doesn't need, so it learns the semantic and dynamic structure of a scene (what moves, how, and in what order) rather than how to repaint it.

The ViT-g variant trades latency for quality: it is the strongest V-JEPA 2 encoder, worth it when representation quality drives your retrieval or classification accuracy more than throughput does. On Mixpeek it serves as a motion-aware video embedding stage, giving an agent a compact vector of what *happens* over a clip, complementary to keyframe/caption features that describe what merely *appears*.

Architecture

Giant Vision Transformer video encoder (ViT-g), the largest V-JEPA 2 checkpoint. The FPC64 variant samples 64 frames and exposes get_vision_features via Transformers; it can also encode a still image by repeating it across the frame dimension. Trained self-supervised by predicting masked spacetime representations in latent space (no pixel decoder), which is the core JEPA distinction from pixel-reconstruction MAEs.

Mixpeek SDK Integration

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so vjepa2-vitg-fpc64-256 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: { "multimodal-embedding": yourVector },
          payload: { source_key: "archive/2026/asset-00412" },
        },
      ],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// multimodal_extractor@v2 runs google/gemini-embedding-2
// (3072-d) over a bucket, with no inference of your own.