gitwtfhub

wtf is vjepa2?

facebookresearch/vjepa2 — explained in plain English

Analysis updated 2026-07-03 · repo last pushed 2026-03-23

★ 4,235PythonAudience · researcherComplexity · 4/5MaintainedLicenseSetup · hard

TL;DR

A self-supervised AI system from Meta that learns to understand video by predicting hidden parts of clips, enabling action recognition, video Q&A, and robot control without labeled data.

Mindmap

mindmap
  root((vjepa2))
    What it does
      Masked video prediction
      Self-supervised learning
      Action recognition
    Capabilities
      Video understanding
      Robot manipulation
      Temporal reasoning
    Model Sizes
      80M parameters
      2B parameters
    Tech Stack
      Python
      PyTorch
      decord
    Audience
      AI researchers
      Robotics engineers

Code map

Detail Auto

An interactive map of this repo's files and how they connect — its source is parsed live in your browser. Click Visualize to build it.

filefunction / class

Why would anyone build with this?

REASON 1

Run video action recognition on your own footage using a pretrained V-JEPA 2 model without labeling any data.

REASON 2

Fine-tune the model for a robotics task to teach a robot arm to pick up objects without task-specific training.

REASON 3

Build a video question-answering feature that can answer temporal questions about what happens in a video clip.

REASON 4

Train linear probes on top of V-JEPA 2 features to benchmark video understanding on your own dataset.

What's in the stack?

PythonPyTorchdecord

How it stacks up

facebookresearch/vjepa2structuredlabs/preswaldfacebookresearch/deit
Stars4,2354,2904,349
LanguagePythonPythonPython
Last pushed2026-03-23—2024-03-15
MaintenanceMaintained—Dormant
Setup difficultyhardeasymoderate
Complexity4/52/53/5
Audienceresearcherdataresearcher

Figures from each repo's GitHub metadata at analysis time.

How do you spin it up?

Difficulty · hard Time to first run · 1h+

Requires GPU and the decord video library, which does not work natively on macOS.

Apache 2.0, use freely including commercially, keep the license notice.

Yoink these prompts

Prompt 1
Using facebookresearch/vjepa2, write Python code to load a pretrained V-JEPA 2 model and classify actions in a local video file.
Prompt 2
Show me how to extract frozen video features from V-JEPA 2 and train a linear probe on a custom action recognition dataset.
Prompt 3
Explain how V-JEPA 2 uses masked video prediction to learn without labeled data and how that helps with robotics tasks.
Prompt 4
How do I run the example inference notebook from facebookresearch/vjepa2 on a GPU server with a sample video?

Frequently asked questions

wtf is vjepa2?

A self-supervised AI system from Meta that learns to understand video by predicting hidden parts of clips, enabling action recognition, video Q&A, and robot control without labeled data.

What language is vjepa2 written in?

Mainly Python. The stack also includes Python, PyTorch, decord.

Is vjepa2 actively maintained?

Maintained — commit in last 6 months (last push 2026-03-23).

What license does vjepa2 use?

Apache 2.0, use freely including commercially, keep the license notice.

How hard is vjepa2 to set up?

Setup difficulty is rated hard, with roughly 1h+ to a first successful run.

Who is vjepa2 for?

Mainly researcher.

View the repo → Decode another repo

This repo across BitVibe Labs

Don't trust strangers blindly. Verify against the repo.

Other repos in this lane