I-JEPA
First model built on LeCun's Joint-Embedding Predictive Architecture, which predicts abstract representations of image regions instead of pixels.
Demonstrated a non-generative, representation-space self-supervised method that is cheaper to train (632M ViT on 16 A100s in under 72 hours) and strong at low-shot ImageNet, the seed of the JEPA / world-model line.
- Date
- Tuesday, 13 June 2023
- Lab
- Meta
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Compute | 632M ViT, 16 A100s, <72 hours Meta says rivals need 2-10x more GPU-hours | company |
Paper was on arXiv in January 2023 (2301.08243); the Meta blog and code release are 2023-06-13.
Sources
This record was checked against its sources on 6 October 2026. How we check