← The openness index

vision · Openness profile

I-JEPA

Meta AI (FAIR) · snapshot 2026-11-07

self-supervised image encoder trained with a Joint-Embedding Predictive Architecture

Tier 2 Restricted open-weight

What the publisher claims: open source, open weights

The one-liner: The image JEPA pioneer, under a non-commercial license.

The six questions

A AccessScore 3 of 3

Weights ungated on Hugging Face.

P PermissionScore 1 of 3

CC-BY-NC 4.0 on both code and weights bars commercial use.

D DeploymentScore 3 of 3

Standard PyTorch, runs on user hardware.

T TransparencyScore 1 of 3

Method paper published; training data (ImageNet) and pipeline not packaged for reproduction.

R ReproducibilityScore 1 of 3

Paper-level detail only; no recipe equivalent to a full release.

G Governance & ParticipationScore 1 of 3

Meta's I-JEPA training and evaluation code is public and takes outside issues and the occasional PR, but the JEPA research agenda and every release are decided inside FAIR.

License

CC-BY-NC 4.0 (code and weights) — license source ↗

What is actually released

Why it matters

Shows the original JEPA release was open-access, not open source — the license moved before the architecture became famous.

Sources

Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.

← The openness indexJoin the newsletter ↗