# facebookresearch/vjepa2

PyTorch code and models for VJEPA2 self-supervised learning from video.

Repository: https://github.com/facebookresearch/vjepa2
Canonical: https://ross.abutalabs.com/products/vjepa2
Language: Python
License: MIT
License Family: permissive
Last push: 2026-03-23T10:13:05+00:00

## Health v2 (maintenance only)
Score: 52/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 73, release rhythm 35, longevity 35
- inputs: {"age_days": 495, "days_push": 163, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 4527, forks 556 (observed 2026-08-28T04:08:52.889168+00:00)

## What it is
Official PyTorch codebase and pretrained models for V-JEPA 2, a self-supervised video encoder trained on internet-scale video, plus V-JEPA 2-AC, an action-conditioned latent world model for robot manipulation. It provides training recipes, model weights, and evaluation code for video understanding, prediction, and planning tasks.

## Use cases
- extract video features for action recognition
- anticipate human actions from video
- train a self-supervised video encoder
- run a robot manipulation policy zero-shot
- fine-tune video models on downstream tasks
- predict future video latents with a world model

## When to choose
- you need state-of-the-art video understanding features in PyTorch
- you want to experiment with self-supervised video pretraining
- you need a video world model for robot planning without task-specific training

## When to avoid
- you need a lightweight production video pipeline rather than research models
- you lack GPU resources for large vision transformers
- you need a general-purpose video editing or playback tool

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, deep-learning, video-processing, computer-vision
- domain: machine-learning, deep-learning, computer-vision, artificial-intelligence, robotics
- platform: python
- tags: self-supervised-learning, video-encoder, world-model, pytorch, research-models, robot-manipulation, je, gpu

## Member repositories
- facebookresearch/vjepa2 (main) score 52

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:08:52.889168+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T18:20:14.851344+00:00, confidence not recorded.
  - readme: https://github.com/facebookresearch/vjepa2 (fetched 2026-08-28T04:08:52.889168+00:00, sha c7c4b4db40af)
- Data as of 2026-08-30T08:39:29.467469+00:00.