Project AURIONAI Use-Cases Research & Innovation
Meta
world model

V-JEPA 2

Meta FAIR latent predictive world model; code and checkpoints released for commercial and research use (licence name not stated).

MetaMetaUSAV-JEPAFamily
Figure from the V-JEPA 2 paper: V-JEPA 2 Overview.
Levelmodel
Classworld model
Parameters1.2B
Contextnot in the notes
Released11 Jun 2025
Opennessopen weights permissive
Runs onnot in the notes
Statusactive

Benchmark Results

vendor-reported
Epic-Kitchens-100 action anticipationmultimodal
anticipationvendor-reported
39.7
02040

Score in R@5. Higher is better.

V-JEPA 2 paperJun 2025 arxiv.org
PerceptionTestmultimodal
8B LLM-aligned scalevendor-reported
84
050100

Score in %. Higher is better.

V-JEPA 2 paperJun 2025 arxiv.org
Something-Something v2multimodal
vendor-reported
77.3
050100

Score in % top-1. Higher is better.

V-JEPA 2 paperJun 2025 arxiv.org
TempCompassmultimodal
8B LLM-aligned scalevendor-reported
76.9
050100

Score in %. Higher is better.

V-JEPA 2 paperJun 2025 arxiv.org
Where this comes from
Every score names who ran it: vendor-reported, maintainer-run or independent. A rank is shown with the population it was ranked in (gap rule 3).

Lineage

finetune
V-JEPA 2 to V-JEPA 2-AC: finetune, high confidenceV-JEPA 2V-JEPA 2Meta, Jun 2025V-JEPA 2-ACV-JEPA 2-ACMeta, Jun 2025
MetaV-JEPA 2-AC is a child (finetune)high confidencepress claim
Where this comes from
Parent and child relations from model cards, Hugging Face base_model fields and the genealogy tables of the research notes. The tree shows two generations each way, 2 models. Confidence and evidence are the notes' own.

R&D Topics

World models (self-supervised video, robotics)exemplar

Identity

Developer
Meta
Family
V-JEPA
Released
11 Jun 2025 sourcesannouncement page
Weights published
–
Record checked
10 Oct 2026
Where this comes from
Developer and dates from the research notes. Release date is the vendor's public announcement; weights date is when open weights were published.

Architecture and Size

Architecture
other Self-supervised joint-embedding predictive video model, pretrained on 1M+ hours of video and images
Total parameters
1.2B sources
Active parameters
–
Checkpoint parameters
–
Note
Over 1M hours of video plus 1M images, then 62 hours of robot data.
Context length
not in the notes
Input
video, image
Output
text
Where this comes from
Parameter counts as the developer states them. Where the checkpoint holds a different total, both are shown (register A16).

Openness and Access

Openness
open weights permissive
Licence
not in the notes sources
Access routes
open weights download
Where this comes from
The licence file governs over registry tags and aggregator labels.

Deployment

Runs on
not in the notes
Footprint
not in the notes
Runtimes
not in the notes
Where this comes from
Deployment tiers and runtimes as named in the notes. Footprint is the notes' figure, not a measurement by AURION.

Indonesia

Indonesian support
not stated
Regional languages
not in the notes
Jakarta hosting
unknown
Languages
not in the notes
Where this comes from
Indonesia fields are carried on every model; unknown is an allowed value.

Industry Use Cases

No industry use case links this model.

Sources

  1. Introducing the V-JEPA 2 world model and new benchmarks for physical reasoning ai.meta.com https://ai.meta.com/blog/v-jepa-2-world-model-benchmarks/
  2. V-JEPA 2: Self-Supervised Video Models Enable Understanding, Prediction and Planning arxiv.org https://arxiv.org/abs/2506.09985

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "V-JEPA 2", https://aurion.semai.id/models/meta-v-jepa-2/, records of October 2026.