Runs as a custom pluginEmbeddingVerified 2026-11-05
V-JEPA 2 VITL
facebook/vjepa2-vitl-fpc64-256
V-JEPA 2 VITL is a video embedding model (MIT). Can run on Mixpeek as a custom plugin.
Mixpeek does not run it today. It can run as a custom plugin: the weights load without remote code.
No demo for V-JEPA 2 VITL yet
A video, image, slideshow or playground goes here once we record one on our own data.
What it is good and bad at
- 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.
- Task
- Embedding
- Input
- Video
- Output
- Vector
- Use cases
- Video understanding, Action recognition
Templates that use it
No template uses this model yet. Browse templates.
Input and output
Input
- video
- frames
- 64
- resolution
- 256x256
Output
- vector
- dims
- 1024
- distance
- cosine
- type
- dense
- note
- patch tokens; pool for a clip vector
How Mixpeek wires it
- Extractors
- None today
- Feature URIs
- None today
- Vector indexes
- None
- Retriever stages
- None
- Combines with
- scene splits -> V-JEPA 2 pooled clip vectors -> feature_search by example clip
- V-JEPA 2 features -> trained probe -> action labels -> attribute filter
How it runs
- Status
- Runs as a custom plugin Not served. Safetensors with transformers VJEPA2Model (no remote code); would run as a GPU batch plugin producing clip vectors; CPU possible but slow at 64 frames.
- Serve apps
- None
- Runtime
- transformers VJEPA2Model + AutoVideoProcessor
- Weights
- safetensors, no remote code
- Parameters
- 326 M
- Size on disk
- 1,304 MB
- Preprocessing
- Sample 64 frames, resize and center-crop, normalize ImageNet mean/std.
- Postprocessing
- Mean-pool patch tokens (or attentive probe); L2-normalize.
- Serving
- batch
- Hardware
- Runs on CPU (slower without a GPU)
Scale to zero
- Minimum replicas
- 0
- Maximum replicas
- 1
- Idle timeout
- 600 s
- Memory
- 6 GB
- CPUs
- 4
- GPUs
- 1
- Cold start, node warm
- Not recorded
- Cold start, new node
- Not recorded
Measured cost
Unmeasured
No committed measurement; plan row 3 measures it in the standalone image.
Our evaluations
Unmeasured
We have not evaluated this model on our own data yet. Publisher results are in the sources below.
Questions
- What is V-JEPA 2 VITL best at?
- Self-supervised video features that capture motion and physical dynamics; strong with attentive probes for action anticipation and classification.
- What are the limits of V-JEPA 2 VITL?
- No text tower: cannot answer text queries without training a probe or alignment; 64-frame clips are heavy.
- Can V-JEPA 2 VITL be used commercially?
- Yes. MIT per the hub tag.
- How does V-JEPA 2 VITL run on Mixpeek?
- Not served. Safetensors with transformers VJEPA2Model (no remote code); would run as a GPU batch plugin producing clip vectors; CPU possible but slow at 64 frames.
- What does V-JEPA 2 VITL output?
- A 1024-dimensional dense vector, compared by cosine similarity.