{
  "id": "vjepa2-vitg-fpc64-256",
  "identity": {
    "family": "vjepa2",
    "display_name": "V-JEPA 2 VITG",
    "hub_id": "facebook/vjepa2-vitg-fpc64-256",
    "api_name": null,
    "revision": null,
    "variants": [
      "facebook/vjepa2-vitg-fpc64-384 (384px)"
    ],
    "license": "Apache-2.0",
    "commercial_use": true,
    "commercial_use_note": "Apache-2.0 per the hub tag.",
    "release_date": "2025-06-11",
    "paper_url": "https://arxiv.org/abs/2506.09985",
    "repo_url": "https://github.com/facebookresearch/vjepa2"
  },
  "task": {
    "category": "embedding",
    "modalities_in": [
      "video"
    ],
    "modalities_out": [
      "vector"
    ],
    "best_at": "Self-supervised video features that capture motion and physical dynamics; strong with attentive probes for action anticipation and classification.",
    "worst_at": "No text tower: cannot answer text queries without training a probe or alignment; 64-frame clips are heavy."
  },
  "io": {
    "input": {
      "video": {
        "frames": 64,
        "resolution": "256x256"
      }
    },
    "output": {
      "vector": {
        "dims": 1408,
        "distance": "cosine",
        "type": "dense",
        "note": "patch tokens; pool for a clip vector"
      }
    }
  },
  "mixpeek_wiring": {
    "extractors": [],
    "feature_uris": [],
    "vector_indexes": [],
    "retriever_stages": [],
    "composes_with": [
      "scene splits -> V-JEPA 2 pooled clip vectors -> feature_search by example clip",
      "V-JEPA 2 features -> trained probe -> action labels -> attribute filter"
    ]
  },
  "run": {
    "status": "plugin",
    "status_reason": "Not served. Safetensors with transformers VJEPA2Model (no remote code); would run as a GPU batch plugin producing clip vectors (too large for CPU scale to zero).",
    "serve_apps": [],
    "runtime": "transformers VJEPA2Model + AutoVideoProcessor",
    "weights_format": "safetensors",
    "trust_remote_code": false,
    "size_on_disk_mb": 4138,
    "params_millions": 1035,
    "preprocessing": "Sample 64 frames, resize and center-crop, normalize ImageNet mean/std.",
    "postprocessing": "Mean-pool patch tokens (or attentive probe); L2-normalize.",
    "serving_shape": "batch",
    "batch_size": null,
    "serve_batch": null,
    "scale_to_zero": {
      "min_replicas": 0,
      "max_replicas": 1,
      "idle_timeout_s": 600,
      "memory_gb": 16,
      "num_cpus": 4,
      "num_gpus": 1,
      "cold_start_s": {
        "warm_node_s": null,
        "new_node_s": null,
        "source": null
      },
      "cpu_fallback": false
    }
  },
  "cost": {
    "unmeasured": true,
    "measured": null,
    "note": "No committed measurement; plan row 3 measures it in the standalone image."
  },
  "evidence": {
    "unmeasured": true,
    "items": []
  },
  "refs": [
    "hf://facebook/vjepa2-vitg-fpc64-256@875c192b7b704b87d1e1d99345769632dd5f739a"
  ],
  "formats": [
    {
      "ref": "hf://facebook/vjepa2-vitg-fpc64-256@875c192b7b704b87d1e1d99345769632dd5f739a",
      "format": "safetensors",
      "file": "model.safetensors",
      "loader": "transformers / open_clip from_pretrained; Mixpeek Hub import (api/models/huggingface.py)",
      "supported_today": true
    },
    {
      "ref": "hf://facebook/vjepa2-vitg-fpc64-256@875c192b7b704b87d1e1d99345769632dd5f739a",
      "format": "pytorch_pt",
      "file": "original/model.pth",
      "loader": "torch.load (pickle); Mixpeek refuses pickle weights (shared/security/torch_pickle.py)",
      "supported_today": false
    }
  ],
  "use_cases": [
    "video-understanding",
    "action-recognition"
  ],
  "recommended": {
    "value": false,
    "for": [],
    "reason": null
  },
  "demos": [],
  "seo": {
    "title": "V-JEPA 2 VITG: embedding model",
    "description": "V-JEPA 2 VITG is a video embedding model (Apache-2.0). Can run on Mixpeek as a custom plugin.",
    "keywords": [
      "V-JEPA 2 VITG",
      "vjepa2",
      "embedding",
      "facebook/vjepa2-vitg-fpc64-256",
      "video-understanding",
      "action-recognition"
    ],
    "faqs": [
      {
        "q": "What is V-JEPA 2 VITG best at?",
        "a": "Self-supervised video features that capture motion and physical dynamics; strong with attentive probes for action anticipation and classification."
      },
      {
        "q": "What are the limits of V-JEPA 2 VITG?",
        "a": "No text tower: cannot answer text queries without training a probe or alignment; 64-frame clips are heavy."
      },
      {
        "q": "Can V-JEPA 2 VITG be used commercially?",
        "a": "Yes. Apache-2.0 per the hub tag."
      },
      {
        "q": "How does V-JEPA 2 VITG run on Mixpeek?",
        "a": "Not served. Safetensors with transformers VJEPA2Model (no remote code); would run as a GPU batch plugin producing clip vectors (too large for CPU scale to zero)."
      },
      {
        "q": "What does V-JEPA 2 VITG output?",
        "a": "A 1408-dimensional dense vector, compared by cosine similarity."
      }
    ]
  },
  "last_verified": "2026-11-05",
  "sources": [
    "https://huggingface.co/facebook/vjepa2-vitg-fpc64-256",
    "https://arxiv.org/abs/2506.09985",
    "https://github.com/facebookresearch/vjepa2",
    "https://huggingface.co/docs/transformers/model_doc/vjepa2"
  ],
  "registry": null,
  "templates": [],
  "usage": {
    "namespaces": null,
    "collections": null,
    "uncounted": true,
    "templates": 0,
    "counted_at": null,
    "source": null
  },
  "facets": {
    "status": "plugin",
    "category": "embedding",
    "family": "vjepa2",
    "license": "Apache-2.0",
    "commercial_use": true,
    "modalities_in": [
      "video"
    ],
    "modalities_out": [
      "vector"
    ],
    "use_cases": [
      "video-understanding",
      "action-recognition"
    ],
    "recommended": false,
    "release_date": "2025-06-11",
    "params_millions": 1035,
    "size_on_disk_mb": 4138,
    "vector_dims": 1408,
    "vector_distance": "cosine",
    "weights_format": "safetensors",
    "formats": [
      "pytorch_pt",
      "safetensors"
    ],
    "loadable_today": true,
    "trust_remote_code": false,
    "cpu_fallback": false,
    "num_gpus": 1,
    "memory_gb": 16,
    "min_replicas": 0,
    "serving_shape": "batch",
    "served": false,
    "ref_schemes": [
      "hf"
    ],
    "has_evidence": false,
    "cost_measured": false,
    "template_count": 0,
    "usage_namespaces": null,
    "usage_collections": null,
    "usage_uncounted": true,
    "has_demo": false,
    "rank_score": 1
  },
  "rank_score": 1
}